class GAN(AlgorithmInterface):
"""
0 is a generated sample
1 is a true sample
maximize D while minimizing G
"""
def __init__(self, model, state_length, action_length, state_bounds,
action_bounds, settings_):
print("Building GAN Model")
super(GAN, self).__init__(model, state_length, action_length,
state_bounds, action_bounds, 0, settings_)
self._noise_mean = 0.0
self._noise_std = 1.0
self._noise_shared = theano.shared(np.zeros(
(self._batch_size, 1), dtype=self.getSettings()['float_type']),
broadcastable=(False, True))
# if settings['action_space_continuous']:
if ('size_of_result_state' in self.getSettings()):
self._experience = ExperienceMemory(
state_length,
action_length,
self.getSettings()['expereince_length'],
continuous_actions=True,
settings=self.getSettings(),
result_state_length=self.getSettings()['size_of_result_state'])
else:
self._experience = ExperienceMemory(
state_length,
action_length,
self.getSettings()['expereince_length'],
continuous_actions=True,
settings=self.getSettings())
self._experience.setStateBounds(copy.deepcopy(self.getStateBounds()))
self._experience.setRewardBounds(copy.deepcopy(self.getRewardBounds()))
self._experience.setActionBounds(copy.deepcopy(self.getActionBounds()))
self._modelTarget = copy.deepcopy(model)
# print ("Initial W " + str(self._w_o.get_value()) )
self._learning_rate = self.getSettings()["fd_learning_rate"]
self._regularization_weight = 1e-5
self._discount_factor = self.getSettings()['discount_factor']
self._rho = self.getSettings()['rho']
self._rms_epsilon = self.getSettings()['rms_epsilon']
self._weight_update_steps = self.getSettings(
)['steps_until_target_network_update']
self._updates = 0
self._decay_weight = self.getSettings()['regularization_weight']
self._critic_regularization_weight = self.getSettings(
)["critic_regularization_weight"]
# self._q_valsA = lasagne.layers.get_output(self._model.getCriticNetwork(), self._model.getStateSymbolicVariable(), deterministic=True)
# self._q_valsA_drop = lasagne.layers.get_output(self._model.getCriticNetwork(), self._model.getStateSymbolicVariable(), deterministic=False)
# self._q_valsNextState = lasagne.layers.get_output(self._model.getCriticNetwork(), self._model.getResultStateSymbolicVariable(), deterministic=True)
# self._q_valsTargetNextState = lasagne.layers.get_output(self._modelTarget.getCriticNetwork(), self._model.getResultStateSymbolicVariable(), deterministic=True)
# self._q_valsTarget = lasagne.layers.get_output(self._modelTarget.getCriticNetwork(), self._model.getStateSymbolicVariable(), deterministic=True)
# self._q_valsTarget_drop = lasagne.layers.get_output(self._modelTarget.getCriticNetwork(), self._model.getStateSymbolicVariable(), deterministic=False)
if ("train_gan_with_gaussian_noise" in self.getSettings()
and (self.getSettings()["train_gan_with_gaussian_noise"])):
inputs_1 = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
self._model.getResultStateSymbolicVariable():
self._model.getResultStates(),
self._model._Noise:
self._noise_shared
}
self._generator_drop = lasagne.layers.get_output(
self._model.getForwardDynamicsNetwork(),
inputs_1,
deterministic=True)
self._generator = lasagne.layers.get_output(
self._model.getForwardDynamicsNetwork(),
inputs_1,
deterministic=True)
else:
inputs_1 = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
self._model.getResultStateSymbolicVariable():
self._model.getResultStates(),
# self._model._Noise: self._noise_shared
}
self._generator = lasagne.layers.get_output(
self._model.getForwardDynamicsNetwork(),
inputs_1,
deterministic=True)
self._generator_drop = lasagne.layers.get_output(
self._model.getForwardDynamicsNetwork(),
inputs_1,
deterministic=False)
# self._q_valsActTarget = lasagne.layers.get_output(self._modelTarget.getForwardDynamicsNetwork(), self._model.getResultStateSymbolicVariable(), deterministic=True)
# self._q_valsActA_drop = lasagne.layers.get_output(self._model.getForwardDynamicsNetwork(), self._model.getStateSymbolicVariable(), deterministic=False)
inputs_ = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
self._model.getResultStateSymbolicVariable():
self._model.getResultStates(),
# self._model.getRewardSymbolicVariable(): self._model.getRewards(),
# self._model._Noise: self._noise_shared
}
self._discriminator = lasagne.layers.get_output(
self._model.getCriticNetwork(), inputs_, deterministic=True)
self._discriminator_drop = lasagne.layers.get_output(
self._model.getCriticNetwork(), inputs_, deterministic=False)
"""
inputs_2 = {
self._modelTarget.getStateSymbolicVariable(): self._model.getResultStates(),
self._modelTarget.getActionSymbolicVariable(): self._model.getActions()
}
"""
self._diff = self._model.getRewardSymbolicVariable(
) - self._discriminator_drop
loss = T.pow(self._diff, 2)
self._loss = T.mean(loss)
self._diff_g = self._model.getResultStateSymbolicVariable(
) - self._generator_drop
loss_g = T.pow(self._diff_g, 2)
self._loss_g = T.mean(loss_g)
# assert len(lasagne.layers.helper.get_all_params(self._l_outA)) == 16
# Need to remove the action layers from these params
self._params = lasagne.layers.helper.get_all_params(
self._model.getCriticNetwork())
print("******Number of Layers is: " + str(
len(
lasagne.layers.helper.get_all_params(
self._model.getCriticNetwork()))))
print("******Number of Action Layers is: " + str(
len(
lasagne.layers.helper.get_all_params(
self._model.getForwardDynamicsNetwork()))))
self._actionParams = lasagne.layers.helper.get_all_params(
self._model.getForwardDynamicsNetwork())
self._givens_ = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
self._model.getResultStateSymbolicVariable():
self._model.getResultStates(),
self._model.getRewardSymbolicVariable():
self._model.getRewards(),
# self._model._Noise: self._noise_shared
}
self._critic_regularization = (
self._critic_regularization_weight *
lasagne.regularization.regularize_network_params(
self._model.getCriticNetwork(), lasagne.regularization.l2))
## MSE update
self._value_grad = T.grad(self._loss + self._critic_regularization,
self._params)
print("Optimizing Value Function with ",
self.getSettings()['optimizer'], " method")
self._updates_ = lasagne.updates.adam(self._value_grad,
self._params,
self._learning_rate,
beta1=0.9,
beta2=0.9,
epsilon=self._rms_epsilon)
if ("train_gan_with_gaussian_noise" in settings_
and (settings_["train_gan_with_gaussian_noise"])):
self._actGivens = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
# self._model.getResultStateSymbolicVariable(): self._model.getResultStates(),
self._model._Noise:
self._noise_shared
}
self._actGivens_MSE = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
self._model.getResultStateSymbolicVariable():
self._model.getResultStates(),
self._model._Noise:
self._noise_shared
}
else:
self._actGivens = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
# self._model.getResultStateSymbolicVariable(): self._model.getResultStates(),
# self._model._Noise: self._noise_shared
}
self._actGivens_MSE = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
# self._model.getResultStateSymbolicVariable(): self._model.getResultStates(),
# self._model._Noise: self._noise_shared
}
self._actor_regularization = (
self._regularization_weight *
lasagne.regularization.regularize_network_params(
self._model.getForwardDynamicsNetwork(),
lasagne.regularization.l2))
## MSE update
self._gen_grad = T.grad(self._loss_g + self._actor_regularization,
self._actionParams)
print("Optimizing Value Function with ",
self.getSettings()['optimizer'], " method")
self._updates_generator = lasagne.updates.adam(
self._gen_grad,
self._actionParams,
self._learning_rate,
beta1=0.9,
beta2=0.9,
epsilon=self._rms_epsilon)
## Some cool stuff to backprop action gradients
self._result_state_grad = T.matrix("Action_Grad")
self._result_state_grad.tag.test_value = np.zeros(
(self._batch_size, self._state_length),
dtype=np.dtype(self.getSettings()['float_type']))
self._result_state_grad_shared = theano.shared(
np.zeros((self._batch_size, self._state_length),
dtype=self.getSettings()['float_type']))
### Maximize wrt q function
self._result_state_mean_grads = T.grad(
cost=None,
wrt=self._actionParams,
known_grads={self._generator: self._result_state_grad_shared}),
print("Action grads: ", self._result_state_mean_grads[0])
## When passing in gradients it needs to be a proper list of gradient expressions
self._result_state_mean_grads = list(self._result_state_mean_grads[0])
# print ("isinstance(self._action_mean_grads, list): ", isinstance(self._action_mean_grads, list))
# print ("Action grads: ", self._action_mean_grads)
self._generatorGRADUpdates = lasagne.updates.adam(
self._result_state_mean_grads,
self._actionParams,
self._learning_rate * 0.1,
beta1=0.9,
beta2=0.9,
epsilon=self._rms_epsilon)
self._givens_grad = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
self._model.getResultStateSymbolicVariable():
self._model.getResultStates(),
# self._model.getRewardSymbolicVariable(): self._model.getRewards(),
}
### Some other stuff to learn a reward function
self._inputs_reward_ = {
self._model.getStateSymbolicVariable(): self._model.getStates(),
self._model.getActionSymbolicVariable(): self._model.getActions(),
}
self._reward = lasagne.layers.get_output(
self._model.getRewardNetwork(),
self._inputs_reward_,
deterministic=True)
self._reward_drop = lasagne.layers.get_output(
self._model.getRewardNetwork(),
self._inputs_reward_,
deterministic=False)
## because rewards are noramlized then scaled by the discount factor to the value stay between -1,1.
self._reward_diff = (self._model.getRewardSymbolicVariable() *
(1.0 /
(1.0 - self.getSettings()['discount_factor']))
) - self._reward_drop
self.__Reward = self._model.getRewardSymbolicVariable()
print("self.__Reward", self.__Reward)
# self._reward_diff = (self._model.getRewardSymbolicVariable()) - self._reward_drop
self._reward_loss_ = T.mean(T.pow(self._reward_diff, 2), axis=1)
self._reward_loss = T.mean(self._reward_loss_)
self._reward_diff_NoDrop = (
self._model.getRewardSymbolicVariable() *
(1.0 /
(1.0 - self.getSettings()['discount_factor']))) - self._reward
# self._reward_diff_NoDrop = (self._model.getRewardSymbolicVariable()) - self._reward
self._reward_loss_NoDrop_ = T.mean(T.pow(self._reward_diff_NoDrop, 2),
axis=1)
self._reward_loss_NoDrop = T.mean(self._reward_loss_NoDrop_)
self._reward_params = lasagne.layers.helper.get_all_params(
self._model.getRewardNetwork())
self._reward_givens_ = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
# self._model.getResultStateSymbolicVariable() : self._model.getResultStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
self._model.getRewardSymbolicVariable():
self._model.getRewards(),
}
self._reward_updates_ = lasagne.updates.adam(
self._reward_loss +
(self._regularization_weight *
lasagne.regularization.regularize_network_params(
self._model.getRewardNetwork(), lasagne.regularization.l2)),
self._reward_params,
self._learning_rate,
beta1=0.9,
beta2=0.999,
epsilon=self._rms_epsilon)
GAN.compile(self)
def compile(self):
self._train = theano.function([], [self._loss, self._discriminator],
updates=self._updates_,
givens=self._givens_)
# self._trainActor = theano.function([], [actLoss, self._q_valsActA], updates=actionUpdates, givens=actGivens)
# self._trainActor = theano.function([], [self._q_func], updates=self._actionUpdates, givens=self._actGivens)
self._trainGenerator = theano.function(
[], [], updates=self._generatorGRADUpdates, givens=self._actGivens)
self._trainGenerator_MSE = theano.function(
[], [],
updates=self._updates_generator,
givens=self._actGivens_MSE)
self._discriminate = theano.function(
[],
self._discriminator,
givens={
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
self._model.getResultStateSymbolicVariable():
self._model.getResultStates(),
})
#self._q_val_Target = theano.function([], self._q_valsB_, givens=self._givens_grad)
if ("train_gan_with_gaussian_noise" in self.getSettings()
and (self.getSettings()["train_gan_with_gaussian_noise"])):
self._generate = theano.function(
[],
self._generator,
givens={
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
# self._model.getResultStateSymbolicVariable(): self._model.getResultStates(),
self._model._Noise:
self._noise_shared
})
else:
self._generate = theano.function(
[],
self._generator,
givens={
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
# self._model.getResultStateSymbolicVariable(): self._model.getResultStates(),
self._model._Noise:
self._noise_shared
})
"""
inputs_ = [
self._model.getStateSymbolicVariable(),
self._model.getRewardSymbolicVariable(),
# ResultState
]
self._bellman_error = theano.function(inputs=inputs_, outputs=self._diff, allow_input_downcast=True)
"""
# self._diffs = theano.function(input=[State])
self._bellman_error = theano.function(
inputs=[],
outputs=self._loss_g,
allow_input_downcast=True,
givens={
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
self._model.getResultStateSymbolicVariable():
self._model.getResultStates(),
self._model._Noise:
self._noise_shared
})
# self._get_action_grad = theano.function([], outputs=lasagne.updates.get_or_compute_grads(T.mean(self._discriminator), [self._model._actionInputVar] + self._params), allow_input_downcast=True, givens=self._givens_grad)
self._get_state_grad = theano.function(
[],
outputs=lasagne.updates.get_or_compute_grads(
T.mean(self._discriminator),
[self._model._stateInputVar] + self._params),
allow_input_downcast=True,
givens=self._givens_grad)
self._get_result_state_grad = theano.function(
[],
outputs=lasagne.updates.get_or_compute_grads(
T.mean(self._discriminator),
[self._model._resultStateInputVar] + self._params),
allow_input_downcast=True,
givens=self._givens_grad)
self._get_action_grad = theano.function(
[],
outputs=T.grad(
cost=None,
wrt=[self._model._actionInputVar] + self._actionParams,
known_grads={self._generator: self._result_state_grad_shared}),
allow_input_downcast=True,
givens=self._actGivens)
# self._get_grad_reward = theano.function([], outputs=lasagne.updates.get_or_compute_grads((self._reward_loss_NoDrop), [lasagne.layers.get_all_layers(self._model.getRewardNetwork())[0].input_var] + self._reward_params), allow_input_downcast=True,
self._get_grad_reward = theano.function(
[],
outputs=lasagne.updates.get_or_compute_grads(
T.mean(self._reward),
[self._model._actionInputVar] + self._reward_params),
allow_input_downcast=True,
givens=self._inputs_reward_)
self._train_reward = theano.function([], [self._reward_loss],
updates=self._reward_updates_,
givens=self._reward_givens_)
self._predict_reward = theano.function([],
self._reward,
givens=self._inputs_reward_)
self._reward_error = theano.function(inputs=[],
outputs=self._reward_diff,
allow_input_downcast=True,
givens=self._reward_givens_)
self._reward_values = theano.function(
inputs=[],
outputs=self.__Reward,
allow_input_downcast=True,
givens={
# self._model.getStateSymbolicVariable() : self._model.getStates(),
# self._model.getResultStateSymbolicVariable() : self._model.getResultStates(),
# self._model.getActionSymbolicVariable(): self._model.getActions(),
self._model.getRewardSymbolicVariable():
self._model.getRewards(),
})
def getStateGrads(self, states, actions=None, alreadyNormed=False):
"""
The states should be normalized
"""
# self.setData(states, actions, rewards, result_states)
if (alreadyNormed == False):
states = norm_state(states, self._state_bounds)
states = np.array(states, dtype=theano.config.floatX)
self._model.setStates(states)
return self._get_state_grad()
def getResultStateGrads(self,
result_states,
actions=None,
alreadyNormed=False):
"""
The states should be normalized
"""
# self.setData(states, actions, rewards, result_states)
if (alreadyNormed == False):
result_states = norm_state(result_states, self._state_bounds)
result_states = np.array(result_states, dtype=theano.config.floatX)
self._model.setResultStates(result_states)
return self._get_result_state_grad()
def setGradTarget(self, grad):
self._result_state_grad_shared.set_value(grad)
def getGrads(self,
states,
actions,
result_states,
v_grad=None,
alreadyNormed=False):
if (alreadyNormed == False):
states = np.array(norm_state(states, self._state_bounds),
dtype=self.getSettings()['float_type'])
actions = np.array(norm_action(actions, self._action_bounds),
dtype=self.getSettings()['float_type'])
result_states = np.array(norm_state(result_states,
self._state_bounds),
dtype=self.getSettings()['float_type'])
# result_states = np.array(result_states, dtype=self.getSettings()['float_type'])
self.setData(states, actions, result_states)
# if (v_grad != None):
self.setGradTarget(v_grad)
return self._get_action_grad()
def getRewardGrads(self, states, actions, alreadyNormed=False):
# states = np.array(states, dtype=self.getSettings()['float_type'])
# actions = np.array(actions, dtype=self.getSettings()['float_type'])
if (alreadyNormed is False):
states = np.array(norm_state(states, self._state_bounds),
dtype=self.getSettings()['float_type'])
actions = np.array(norm_action(actions, self._action_bounds),
dtype=self.getSettings()['float_type'])
# rewards = np.array(norm_state(rewards, self._reward_bounds), dtype=self.getSettings()['float_type'])
self.setData(states, actions)
return self._get_grad_reward()
def getNetworkParameters(self):
params = []
params.append(
lasagne.layers.helper.get_all_param_values(
self._model.getCriticNetwork()))
params.append(
lasagne.layers.helper.get_all_param_values(
self._model.getForwardDynamicsNetwork()))
params.append(
lasagne.layers.helper.get_all_param_values(
self._model.getRewardNetwork()))
params.append(
lasagne.layers.helper.get_all_param_values(
self._modelTarget.getCriticNetwork()))
params.append(
lasagne.layers.helper.get_all_param_values(
self._modelTarget.getForwardDynamicsNetwork()))
params.append(
lasagne.layers.helper.get_all_param_values(
self._modelTarget.getRewardNetwork()))
return params
def setNetworkParameters(self, params):
lasagne.layers.helper.set_all_param_values(
self._model.getCriticNetwork(), params[0])
lasagne.layers.helper.set_all_param_values(
self._model.getForwardDynamicsNetwork(), params[1])
lasagne.layers.helper.set_all_param_values(
self._model.getRewardNetwork(), params[2])
lasagne.layers.helper.set_all_param_values(
self._modelTarget.getCriticNetwork(), params[3])
lasagne.layers.helper.set_all_param_values(
self._modelTarget.getForwardDynamicsNetwork(), params[4])
lasagne.layers.helper.set_all_param_values(
self._modelTarget.getRewardNetwork(), params[5])
def setData(self, states, actions, result_states=None, rewards=None):
self._model.setStates(states)
self._model.setActions(actions)
if not (result_states is None):
self._model.setResultStates(result_states)
if not (rewards is None):
self._model.setRewards(rewards)
noise = np.random.normal(self._noise_mean,
self._noise_std,
size=(states.shape[0], 1))
self._noise_shared.set_value(noise)
# noise = np.zeros((states.shape[0],1))
# self._noise_shared.set_value(noise)
def trainCritic(self, states, actions, result_states, rewards):
self.setData(states, actions, result_states, rewards)
noise = np.random.normal(self._noise_mean,
self._noise_std,
size=(states.shape[0], 1))
# print ("Shapes: ", states.shape, actions.shape, rewards.shape, result_states.shape, falls.shape, noise.shape)
self._noise_shared.set_value(noise)
self._updates += 1
## Compute actions for TargetNet
generated_samples = self._generate()
### Put generated samples in memory
for i in range(generated_samples.shape[0]):
next_state__ = scale_state(generated_samples[i],
self._state_bounds)
tup = ([states[i]], [actions[i]], [next_state__], [rewards[i]],
[0], [0], [0])
self._experience.insertTuple(tup)
tmp_result_states = copy.deepcopy(result_states)
tmp_rewards = copy.deepcopy(rewards)
## Pull out a batch of generated samples
states__, actions__, generated_samples, rewards__, falls__, G_ts__, exp_actions__ = self._experience.get_batch(
min(states.shape[0], self._experience.samples()))
"""
print("generated_samples: ", generated_samples.shape)
print("tmp_result_states: ", tmp_result_states.shape)
print("tmp_rewards: ", tmp_rewards.shape)
print("states: ", states.shape)
print("actions: ", actions.shape)
"""
## replace half of the samples with generated ones...
for i in range(int(states.shape[0] / 2)):
tmp_result_states[i] = generated_samples[i]
tmp_rewards[i] = [0]
# print("Discriminator targets: ", tmp_rewards)
self.setData(states, actions, tmp_result_states, tmp_rewards)
loss, _ = self._train()
# print("Discriminator loss: ", loss)
return loss
def trainActor(self, states, actions, result_states, rewards):
self.setData(states, actions, result_states, rewards)
# self._noise_shared.set_value(np.random.normal(self._noise_mean,self._noise_std, size=(states.shape[0],1)))
## Add MSE term
if ('train_gan_mse' in self.getSettings()
and (self.getSettings()['train_gan_mse'] == False)):
pass
else:
self._trainGenerator_MSE()
# print("Policy mean: ", np.mean(self._q_action(), axis=0))
loss = 0
# print("******** Not learning actor right now *****")
# return loss
generated_samples = self.predict_batch(states, actions)
result_state_grads = self.getResultStateGrads(generated_samples,
actions,
alreadyNormed=True)[0]
discriminator_value = self._discriminate()
"""
From DEEP REINFORCEMENT LEARNING IN PARAMETERIZED ACTION SPACE
Hausknecht, Matthew and Stone, Peter
actions.shape == result_state_grads.shape
"""
use_parameter_grad_inversion = True
if (use_parameter_grad_inversion):
for i in range(result_state_grads.shape[0]):
for j in range(result_state_grads.shape[1]):
if (result_state_grads[i, j] > 0):
inversion = (1.0 - generated_samples[i, j]) / 2.0
else:
inversion = (generated_samples[i, j] - (-1.0)) / 2.0
result_state_grads[i,
j] = result_state_grads[i,
j] * inversion
if (self.getSettings()["print_levels"][self.getSettings(
)["print_level"]] >= self.getSettings()["print_levels"]['debug']):
print("Policy mean: ", np.mean(self._generate(), axis=0))
print("Mean action grad: ", np.mean(result_state_grads, axis=0),
" std ", np.std(result_state_grads, axis=0))
## Set data for gradient
self._model.setResultStates(result_states)
self._modelTarget.setResultStates(result_states)
# self._noise_shared.set_value(np.random.normal(self._noise_mean,self._noise_std, size=(states.shape[0],1)))
error_MSE = self._bellman_error()
## Why the -1.0??
## Because the SGD method is always performing MINIMIZATION!!
self._result_state_grad_shared.set_value(-1.0 * result_state_grads)
self._trainGenerator()
# self._noise_shared.set_value(np.random.normal(self._noise_mean,self._noise_std, size=(states.shape[0],1)))
error_MSE = self._bellman_error()
return (np.mean(discriminator_value), error_MSE)
def train(self, states, actions, result_states, rewards):
loss = self.trainCritic(states, actions, result_states, rewards)
# loss = 0
lossActor = self.trainActor(states, actions, result_states, rewards)
if (self.getSettings()['train_reward_predictor']):
# print ("self._reward_bounds: ", self._reward_bounds)
# print( "Rewards, predicted_reward, difference, model diff, model rewards: ", np.concatenate((rewards, self._predict_reward(), self._predict_reward() - rewards, self._reward_error(), self._reward_values()), axis=1))
self.setData(states, actions, result_states, rewards)
lossReward = self._train_reward()
if (self.getSettings()["print_levels"][self.getSettings(
)["print_level"]] >= self.getSettings()["print_levels"]['train']):
print("Loss Reward: ", lossReward)
return (loss, lossActor)
def predict(self, state, deterministic_=True):
pass
def predict_batch(self, states, deterministic_=True):
pass
def predict(self, state, action):
# states = np.zeros((self._batch_size, self._self._state_length), dtype=theano.config.floatX)
# states[0, ...] = state
state = np.array(norm_state(state, self._state_bounds),
dtype=self.getSettings()['float_type'])
# print ("fd state: ", state)
action = np.array(norm_action(action, self._action_bounds),
dtype=self.getSettings()['float_type'])
# self._model.setStates(state)
# self._model.setActions(action)
self.setData(state, action)
# self._noise_shared.set_value(np.random.normal(self._noise_mean,self._noise_std, size=(1,1)))
# print ("State bounds: ", self._state_bounds)
# print ("gen output: ", self._generate()[0])
state_ = scale_state(self._generate(), self._state_bounds)
# print( "self._state_bounds: ", self._state_bounds)
# print ("scaled output: ", state_)
return state_
def predict_batch(self, states, actions):
## These input should already be normalized.
# self._model.setStates(states)
# self._model.setActions(actions)
self.setData(states, actions)
# self._noise_shared.set_value(np.random.normal(self._noise_mean,self._noise_std, size=(states.shape[0],1)))
# print ("State bounds: ", self._state_bounds)
# print ("fd output: ", self._forwardDynamics()[0])
# state_ = scale_state(self._generate(), self._state_bounds)
state_ = self._generate()
return state_
def q_value(self, state):
"""
For returning a vector of q values, state should NOT be normalized
"""
# states = np.zeros((self._batch_size, self._state_length), dtype=theano.config.floatX)
# states[0, ...] = state
state = norm_state(state, self._state_bounds)
state = np.array(state, dtype=theano.config.floatX)
self._model.setStates(state)
self._modelTarget.setStates(state)
action = self._q_action()
self._model.setActions(action)
self._modelTarget.setActions(action)
return scale_reward(self._discriminate(), self.getRewardBounds()) * (
1.0 / (1.0 - self.getSettings()['discount_factor']))
# return self._q_valTarget()[0]
# return self._q_val()[0]
def q_value(self, state, action, next_state):
"""
For returning a vector of q values, state should NOT be normalized
"""
# states = np.zeros((self._batch_size, self._state_length), dtype=theano.config.floatX)
# states[0, ...] = state
state = norm_state(state, self._state_bounds)
state = np.array(state, dtype=theano.config.floatX)
self._model.setStates(state)
self._modelTarget.setStates(state)
# action = self._q_action()
action = norm_state(action, self.getActionBounds())
self._model.setActions(action)
self._modelTarget.setActions(action)
nextState = norm_state(next_state, self.getStateBounds())
# nextState = np.reshape(nextState, (1,20))
self._model.setResultStates(nextState)
self._modelTarget.setResultStates(nextState)
# return scale_reward(self._discriminate(), self.getRewardBounds())[0] * (1.0 / (1.0- self.getSettings()['discount_factor']))
return self._discriminate()
# return self._q_valTarget()[0]
# return self._q_val()[0]
def q_values(self, state):
"""
For returning a vector of q values, state should already be normalized
"""
state = norm_state(state, self._state_bounds)
state = np.array(state, dtype=theano.config.floatX)
self._model.setStates(state)
self._modelTarget.setStates(state)
action = self._q_action()
self._model.setActions(action)
self._modelTarget.setActions(action)
return scale_reward(self._q_val(), self.getRewardBounds()) * (
1.0 / (1.0 - self.getSettings()['discount_factor']))
# return self._q_valTarget()
# return self._q_val()
def predict_std(self, state, deterministic_=True):
"""
This does nothing for a GAN...
"""
# states = np.zeros((self._batch_size, self._state_length), dtype=theano.config.floatX)
# states[0, ...] = state
action_std = np.array([0] * len(self._action_bounds))
# np.zeros((state.shape[0], len(self._action_bounds)))
# else:
# action_ = scale_action(self._q_action()[0], self._action_bounds)
# action_ = q_valsActA[0]
return action_std
def predict_reward(self, state, action):
# states = np.zeros((self._batch_size, self._self._state_length), dtype=theano.config.floatX)
# states[0, ...] = state
state = np.array(norm_state(state, self._state_bounds),
dtype=self.getSettings()['float_type'])
action = np.array(norm_action(action, self._action_bounds),
dtype=self.getSettings()['float_type'])
self._model.setStates(state)
self._model.setActions(action)
predicted_reward = self._predict_reward()
reward_ = scale_reward(predicted_reward, self.getRewardBounds(
)) # * (1.0 / (1.0- self.getSettings()['discount_factor']))
# reward_ = scale_reward(predicted_reward, self.getRewardBounds())[0] * (1.0 / (1.0- self.getSettings()['discount_factor']))
# reward_ = scale_state(predicted_reward, self._reward_bounds)
# print ("reward, predicted reward: ", reward_, predicted_reward)
return reward_
def predict_reward_batch(self, states, actions):
# states = np.zeros((self._batch_size, self._self._state_length), dtype=theano.config.floatX)
# states[0, ...] = state
# state = np.array(norm_state(state, self._state_bounds), dtype=self.getSettings()['float_type'])
# action = np.array(norm_action(action, self._action_bounds), dtype=self.getSettings()['float_type'])
self._model.setStates(states)
self._model.setActions(actions)
predicted_reward = self._predict_reward()
# reward_ = scale_reward(predicted_reward, self.getRewardBounds())[0] # * (1.0 / (1.0- self.getSettings()['discount_factor']))
# reward_ = scale_reward(predicted_reward, self.getRewardBounds())[0] * (1.0 / (1.0- self.getSettings()['discount_factor']))
# reward_ = scale_state(predicted_reward, self._reward_bounds)
# print ("reward, predicted reward: ", reward_, predicted_reward)
return predicted_reward
def bellman_error(self, states, actions, result_states, rewards):
self.setData(states, actions, result_states, rewards)
return self._bellman_error()
def reward_error(self, states, actions, result_states, rewards):
# rewards = rewards * (1.0/(1.0-self.getSettings()['discount_factor'])) # scale rewards
self.setData(states, actions, result_states, rewards)
return self._reward_error()
def setStateBounds(self, state_bounds):
super(GAN, self).setStateBounds(state_bounds)
"""
print ("")
print("Setting GAN state bounds: ", state_bounds)
print("self.getStateBounds(): ", self.getStateBounds())
print ("")
"""
self._experience.setStateBounds(copy.deepcopy(self.getStateBounds()))
def setActionBounds(self, action_bounds):
super(GAN, self).setActionBounds(action_bounds)
self._experience.setActionBounds(copy.deepcopy(self.getActionBounds()))
def setRewardBounds(self, reward_bounds):
super(GAN, self).setRewardBounds(reward_bounds)
self._experience.setRewardBounds(copy.deepcopy(self.getRewardBounds()))
old_states = states
# print states
actions = np.array(list(map(f, states)))
actionsNoNoise = np.array(list(map(f, states)))
# states2 = np.transpose(np.repeat([states], 2, axis=0))
# print states2
model = createForwardDynamicsModel(settings, state_bounds, action_bounds,
None, None, None)
experience = ExperienceMemory(len(state_bounds[0]),
len(action_bounds[0]),
experience_length,
continuous_actions=True,
settings=settings)
experience.setStateBounds(state_bounds)
experience.setRewardBounds(reward_bounds)
experience.setActionBounds(action_bounds)
experience.setSettings(settings)
arr = list(range(experience_length))
random.shuffle(arr)
num_samples_to_keep = 300
given_actions = []
given_states = []
for i in range(num_samples_to_keep):
a = actions[arr[i]]
action_ = np.array([a])
given_actions.append(action_)
state_ = np.array([states[arr[i]]])
given_states.append(state_)
# print "Action: " + str([actions[i]])
def fitModelToData(settingsFileName):
"""
State is the input state and Action is the desired output (y).
"""
# from model.ModelUtil import *
file = open(settingsFileName)
settings = json.load(file)
print ("Settings: " + str(json.dumps(settings)))
file.close()
import os
os.environ['THEANO_FLAGS'] = "mode=FAST_RUN,device="+settings['training_processor_type']+",floatX="+settings['float_type']
from ModelEvaluation import SimWorker, evalModelParrallel, collectExperience, simEpoch, evalModel
from model.ModelUtil import validBounds
from model.LearningAgent import LearningAgent, LearningWorker
from util.SimulationUtil import validateSettings, createEnvironment, createRLAgent, createActor
from util.SimulationUtil import getDataDirectory, createForwardDynamicsModel, createSampler
from util.ExperienceMemory import ExperienceMemory
from RLVisualize import RLVisualize
from NNVisualize import NNVisualize
from sim.PendulumEnvState import PendulumEnvState
from sim.PendulumEnv import PendulumEnv
from sim.BallGame2DEnv import BallGame2DEnv
import time
settings = validateSettings(settings)
train_forward_dynamics=True
model_type= settings["model_type"]
directory= getDataDirectory(settings)
discrete_actions = np.array(settings['discrete_actions'])
num_actions= discrete_actions.shape[0] # number of rows
rounds = settings["rounds"]
epochs = settings["epochs"]
epsilon = settings["epsilon"]
discount_factor=settings["discount_factor"]
reward_bounds=np.array(settings["reward_bounds"])
batch_size=settings["batch_size"]
train_on_validation_set=settings["train_on_validation_set"]
state_bounds = np.array(settings['state_bounds'])
discrete_actions = np.array(settings['discrete_actions'])
print ("Sim config file name: ", str(settings["sim_config_file"]))
action_space_continuous=settings['action_space_continuous']
if action_space_continuous:
action_bounds = np.array(settings["action_bounds"], dtype=float)
if action_space_continuous:
experience = ExperienceMemory(len(state_bounds[0]), len(action_bounds[0]), settings['expereince_length'], continuous_actions=True, settings=settings)
else:
experience = ExperienceMemory(len(state_bounds[0]), 1, settings['expereince_length'])
file_name=directory+getAgentName()+"expBufferInit.hdf5"
experience.loadFromFile(file_name)
state_bounds = experience._state_bounds
action_bounds = experience._action_bounds
reward_bounds = experience._reward_bounds
output_experience_queue = multiprocessing.Queue(settings['queue_size_limit'])
mgr = multiprocessing.Manager()
namespace = mgr.Namespace()
learning_workers = []
for process in range(1):
# this is the process that selects which game to play
agent = LearningAgent(n_in=len(state_bounds[0]), n_out=len(action_bounds[0]), state_bounds=state_bounds,
action_bounds=action_bounds, reward_bound=reward_bounds, settings_=settings)
agent.setSettings(settings)
lw = LearningWorker(output_experience_queue, agent, namespace)
learning_workers.append(lw)
masterAgent = agent
masterAgent.setExperience(experience)
if action_space_continuous:
model = createRLAgent(settings['agent_name'], state_bounds, action_bounds, reward_bounds, settings)
else:
model = createRLAgent(settings['agent_name'], state_bounds, discrete_actions, reward_bounds, settings)
if ( not settings['load_saved_model'] ):
model.setStateBounds(state_bounds)
model.setActionBounds(action_bounds)
model.setRewardBounds(reward_bounds)
else: # continuation learning
experience.setStateBounds(model.getStateBounds())
experience.setRewardBounds(model.getRewardBounds())
experience.setActionBounds(model.getActionBounds())
if (settings['train_forward_dynamics']):
print ("Created forward dynamics network")
forwardDynamicsModel = createForwardDynamicsModel(settings, state_bounds, action_bounds, None, None)
masterAgent.setForwardDynamics(forwardDynamicsModel)
forwardDynamicsModel.setActor(actor)
forwardDynamicsModel.init(len(state_bounds[0]), len(action_bounds[0]), state_bounds, action_bounds, actor, None, settings)
namespace.forwardNN = masterAgent.getForwardDynamics().getNetworkParameters()
namespace.forwardDynamicsModel = forwardDynamicsModel
## Now everything related to the exp memory needs to be updated
bellman_errors=[]
masterAgent.setPolicy(model)
namespace.agentPoly = masterAgent.getPolicy().getNetworkParameters()
namespace.model = model
if (settings['visualize_learning']):
rlv = NNVisualize(title=str(directory), settings=settings)
rlv.setInteractive()
rlv.init()
if (settings['debug_critic']):
criticLosses = []
criticRegularizationCosts = []
if (settings['visualize_learning']):
critic_loss_viz = NNVisualize(title=str("Critic Loss") + " with " + str(settings["model_type"]))
critic_loss_viz.setInteractive()
critic_loss_viz.init()
critic_regularization_viz = NNVisualize(title=str("Critic Regularization Cost") + " with " + str(settings["model_type"]))
critic_regularization_viz.setInteractive()
critic_regularization_viz.init()
if (settings['debug_actor']):
actorLosses = []
actorRegularizationCosts = []
if (settings['visualize_learning']):
actor_loss_viz = NNVisualize(title=str("Actor Loss") + " with " + str(settings["model_type"]))
actor_loss_viz.setInteractive()
actor_loss_viz.init()
actor_regularization_viz = NNVisualize(title=str("Actor Regularization Cost") + " with " + str(settings["model_type"]))
actor_regularization_viz.setInteractive()
actor_regularization_viz.init()
trainData = {}
trainData["mean_reward"]=[]
trainData["std_reward"]=[]
trainData["mean_bellman_error"]=[]
trainData["std_bellman_error"]=[]
trainData["mean_discount_error"]=[]
trainData["std_discount_error"]=[]
trainData["mean_forward_dynamics_loss"]=[]
trainData["std_forward_dynamics_loss"]=[]
trainData["mean_eval"]=[]
trainData["std_eval"]=[]
trainData["mean_critic_loss"]=[]
trainData["std_critic_loss"]=[]
trainData["mean_critic_regularization_cost"]=[]
trainData["std_critic_regularization_cost"]=[]
trainData["mean_actor_loss"]=[]
trainData["std_actor_loss"]=[]
trainData["mean_actor_regularization_cost"]=[]
trainData["std_actor_regularization_cost"]=[]
best_dynamicsLosses=1000000
_states, _actions, _result_states, _rewards, _falls, _G_ts = experience.get_batch(batch_size)
for round_ in range(rounds):
t0 = time.time()
__states, __actions, __result_states, __rewards, __falls, __G_ts = experience.get_batch(100)
for i in range(1):
masterAgent.train(_states=__states, _actions=__actions, _rewards=__rewards, _result_states=__result_states, _falls=__falls)
t1 = time.time()
time_taken = t1 - t0
if masterAgent.getExperience().samples() > batch_size:
states, actions, result_states, rewards, falls, G_ts = masterAgent.getExperience().get_batch(batch_size)
print ("Batch size: " + str(batch_size))
error = masterAgent.bellman_error(states, actions, rewards, result_states, falls)
bellman_errors.append(error)
if (settings['debug_critic']):
loss__ = masterAgent.getPolicy()._get_critic_loss() # uses previous call batch data
criticLosses.append(loss__)
regularizationCost__ = masterAgent.getPolicy()._get_critic_regularization()
criticRegularizationCosts.append(regularizationCost__)
if (settings['debug_actor']):
loss__ = masterAgent.getPolicy()._get_actor_loss() # uses previous call batch data
actorLosses.append(loss__)
regularizationCost__ = masterAgent.getPolicy()._get_actor_regularization()
actorRegularizationCosts.append(regularizationCost__)
if not all(np.isfinite(error)):
print ("States: " + str(states) + " ResultsStates: " + str(result_states) + " Rewards: " + str(rewards) + " Actions: " + str(actions) + " Falls: ", str(falls))
print ("Bellman Error is Nan: " + str(error) + str(np.isfinite(error)))
sys.exit()
error = np.mean(np.fabs(error))
if error > 10000:
print ("Error to big: ")
print (states, actions, rewards, result_states)
if (settings['train_forward_dynamics']):
dynamicsLoss = masterAgent.getForwardDynamics().bellman_error(states, actions, result_states, rewards)
dynamicsLoss = np.mean(np.fabs(dynamicsLoss))
dynamicsLosses.append(dynamicsLoss)
if (settings['train_forward_dynamics']):
print ("Round: " + str(round_) + " bellman error: " + str(error) + " ForwardPredictionLoss: " + str(dynamicsLoss) + " in " + str(time_taken) + " seconds")
else:
print ("Round: " + str(round_) + " bellman error: " + str(error) + " in " + str(time_taken) + " seconds")
print ("Master agent experience size: " + str(masterAgent.getExperience().samples()))
trainData["mean_bellman_error"].append(np.mean(np.fabs(bellman_errors)))
trainData["std_bellman_error"].append(np.std(bellman_errors))
if (settings['visualize_learning']):
rlv.updateLoss(np.array(trainData["mean_bellman_error"]), np.array(trainData["std_bellman_error"]))
rlv.redraw()
rlv.setInteractiveOff()
rlv.saveVisual(directory+"trainingGraphNN")
rlv.setInteractive()
# print "Error: " + str(error)
if (settings['debug_critic']):
mean_criticLosses = np.mean(criticLosses)
std_criticLosses = np.std(criticLosses)
trainData["mean_critic_loss"].append(mean_criticLosses)
trainData["std_critic_loss"].append(std_criticLosses)
criticLosses = []
if (settings['visualize_learning']):
critic_loss_viz.updateLoss(np.array(trainData["mean_critic_loss"]), np.array(trainData["std_critic_loss"]))
critic_loss_viz.redraw()
critic_loss_viz.setInteractiveOff()
critic_loss_viz.saveVisual(directory+"criticLossGraph")
critic_loss_viz.setInteractive()
mean_criticRegularizationCosts = np.mean(criticRegularizationCosts)
std_criticRegularizationCosts = np.std(criticRegularizationCosts)
trainData["mean_critic_regularization_cost"].append(mean_criticRegularizationCosts)
trainData["std_critic_regularization_cost"].append(std_criticRegularizationCosts)
criticRegularizationCosts = []
if (settings['visualize_learning']):
critic_regularization_viz.updateLoss(np.array(trainData["mean_critic_regularization_cost"]), np.array(trainData["std_critic_regularization_cost"]))
critic_regularization_viz.redraw()
critic_regularization_viz.setInteractiveOff()
critic_regularization_viz.saveVisual(directory+"criticRegularizationGraph")
critic_regularization_viz.setInteractive()
if (settings['debug_actor']):
mean_actorLosses = np.mean(actorLosses)
std_actorLosses = np.std(actorLosses)
trainData["mean_actor_loss"].append(mean_actorLosses)
trainData["std_actor_loss"].append(std_actorLosses)
actorLosses = []
if (settings['visualize_learning']):
actor_loss_viz.updateLoss(np.array(trainData["mean_actor_loss"]), np.array(trainData["std_actor_loss"]))
actor_loss_viz.redraw()
actor_loss_viz.setInteractiveOff()
actor_loss_viz.saveVisual(directory+"actorLossGraph")
actor_loss_viz.setInteractive()
mean_actorRegularizationCosts = np.mean(actorRegularizationCosts)
std_actorRegularizationCosts = np.std(actorRegularizationCosts)
trainData["mean_actor_regularization_cost"].append(mean_actorRegularizationCosts)
trainData["std_actor_regularization_cost"].append(std_actorRegularizationCosts)
actorRegularizationCosts = []
if (settings['visualize_learning']):
actor_regularization_viz.updateLoss(np.array(trainData["mean_actor_regularization_cost"]), np.array(trainData["std_actor_regularization_cost"]))
actor_regularization_viz.redraw()
actor_regularization_viz.setInteractiveOff()
actor_regularization_viz.saveVisual(directory+"actorRegularizationGraph")
actor_regularization_viz.setInteractive()
def trainModelParallel(inputData):
settingsFileName = inputData[0]
settings = inputData[1]
np.random.seed(int(settings['random_seed']))
import os
if ('THEANO_FLAGS' in os.environ):
os.environ['THEANO_FLAGS'] = os.environ[
'THEANO_FLAGS'] + "mode=FAST_RUN,device=" + settings[
'training_processor_type'] + ",floatX=" + settings['float_type']
else:
os.environ['THEANO_FLAGS'] = "mode=FAST_RUN,device=" + settings[
'training_processor_type'] + ",floatX=" + settings['float_type']
import keras.backend
keras.backend.set_floatx(settings['float_type'])
print("K.floatx()", keras.backend.floatx())
from ModelEvaluation import SimWorker, evalModelParrallel, collectExperience, simEpoch, evalModel, simModelParrallel
from model.ModelUtil import validBounds, fixBounds, anneal_value
from model.LearningAgent import LearningAgent, LearningWorker
from util.SimulationUtil import validateSettings
from util.SimulationUtil import createEnvironment
from util.SimulationUtil import createRLAgent
from util.SimulationUtil import createActor, getAgentName
from util.SimulationUtil import getDataDirectory, createForwardDynamicsModel, createSampler
from util.ExperienceMemory import ExperienceMemory
from RLVisualize import RLVisualize
from NNVisualize import NNVisualize
#from sim.PendulumEnvState import PendulumEnvState
#from sim.PendulumEnv import PendulumEnv
#from sim.BallGame2DEnv import BallGame2DEnv
settings = validateSettings(settings)
model_type = settings["model_type"]
directory = getDataDirectory(settings)
if not os.path.exists(directory):
os.makedirs(directory)
# copy settings file
out_file_name = directory + os.path.basename(settingsFileName)
print("Saving settings file with data: ", out_file_name)
out_file = open(out_file_name, 'w')
out_file.write(json.dumps(settings, indent=4))
out_file.close()
### Try and save algorithm and model files for reference
if "." in settings['model_type']:
### convert . to / and copy file over
file_name = settings['model_type']
k = file_name.rfind(".")
file_name = file_name[:k]
file_name_read = file_name.replace(".", "/")
file_name_read = file_name_read + ".py"
print("model file name:", file_name)
print("os.path.basename(file_name): ", os.path.basename(file_name))
file = open(file_name_read, 'r')
out_file = open(directory + file_name + ".py", 'w')
out_file.write(file.read())
file.close()
out_file.close()
if "." in settings['agent_name']:
### convert . to / and copy file over
file_name = settings['agent_name']
k = file_name.rfind(".")
file_name = file_name[:k]
file_name_read = file_name.replace(".", "/")
file_name_read = file_name_read + ".py"
print("model file name:", file_name)
print("os.path.basename(file_name): ", os.path.basename(file_name))
file = open(file_name_read, 'r')
out_file = open(directory + file_name + ".py", 'w')
out_file.write(file.read())
file.close()
out_file.close()
if (settings['train_forward_dynamics']):
if "." in settings['forward_dynamics_model_type']:
### convert . to / and copy file over
file_name = settings['forward_dynamics_model_type']
k = file_name.rfind(".")
file_name = file_name[:k]
file_name_read = file_name.replace(".", "/")
file_name_read = file_name_read + ".py"
print("model file name:", file_name)
print("os.path.basename(file_name): ", os.path.basename(file_name))
file = open(file_name_read, 'r')
out_file = open(directory + file_name + ".py", 'w')
out_file.write(file.read())
file.close()
out_file.close()
rounds = settings["rounds"]
epochs = settings["epochs"]
epsilon = settings["epsilon"]
discount_factor = settings["discount_factor"]
reward_bounds = np.array(settings["reward_bounds"])
batch_size = settings["batch_size"]
train_on_validation_set = settings["train_on_validation_set"]
state_bounds = np.array(settings['state_bounds'])
discrete_actions = np.array(settings['discrete_actions']) #9*6
num_actions = discrete_actions.shape[0] # number of rows
print("Sim config file name: " + str(settings["sim_config_file"]))
action_space_continuous = settings['action_space_continuous']
if (settings['num_available_threads'] == 1):
input_anchor_queue = multiprocessing.Queue(
settings['queue_size_limit'])
input_anchor_queue_eval = multiprocessing.Queue(
settings['queue_size_limit'])
output_experience_queue = multiprocessing.Queue(
settings['queue_size_limit'])
eval_episode_data_queue = multiprocessing.Queue(
settings['queue_size_limit'])
else:
input_anchor_queue = multiprocessing.Queue(settings['epochs'])
input_anchor_queue_eval = multiprocessing.Queue(settings['epochs'])
output_experience_queue = multiprocessing.Queue(
settings['queue_size_limit'])
eval_episode_data_queue = multiprocessing.Queue(
settings['eval_epochs'])
if (settings['on_policy']): ## So that off policy agent does not learn
output_experience_queue = None
sim_work_queues = []
action_space_continuous = settings['action_space_continuous']
if action_space_continuous:
action_bounds = np.array(settings["action_bounds"], dtype=float)
### Using a wrapper for the type of actor now
actor = createActor(settings['environment_type'], settings, None)
exp_val = None
if (not validBounds(action_bounds)):
# Check that the action bounds are spcified correctly
print("Action bounds invalid: ", action_bounds)
sys.exit()
if (not validBounds(state_bounds)):
# Probably did not collect enough bootstrapping samples to get good state bounds.
print("State bounds invalid: ", state_bounds)
state_bounds = fixBounds(np.array(state_bounds))
bound_fixed = validBounds(state_bounds)
print("State bounds fixed: ", bound_fixed)
sys.exit()
if (not validBounds(reward_bounds)):
print("Reward bounds invalid: ", reward_bounds)
sys.exit()
if settings['action_space_continuous']:
experience = ExperienceMemory(len(state_bounds[0]),
len(action_bounds[0]),
settings['expereince_length'],
continuous_actions=True,
settings=settings)
else:
experience = ExperienceMemory(len(state_bounds[0]), 1,
settings['expereince_length'])
experience.setSettings(settings)
if settings['visualize_learning']:
title = settings['agent_name']
k = title.rfind(".") + 1
if (k > len(title)): ## name does not contain a .
k = 0
title = title[k:]
rlv = RLVisualize(title=title + " agent on " +
str(settings['environment_type']),
settings=settings)
rlv.setInteractive()
rlv.init()
if (settings['train_forward_dynamics']):
if settings['visualize_learning']:
title = settings['forward_dynamics_model_type']
k = title.rfind(".") + 1
if (k > len(title)): ## name does not contain a .
k = 0
title = title[k:]
nlv = NNVisualize(title=str("Dynamics Model") + " with " + title,
settings=settings)
nlv.setInteractive()
nlv.init()
if (settings['train_reward_predictor']):
if settings['visualize_learning']:
title = settings['forward_dynamics_model_type']
k = title.rfind(".") + 1
if (k > len(title)): ## name does not contain a .
k = 0
title = title[k:]
rewardlv = NNVisualize(title=str("Reward Model") + " with " +
title,
settings=settings)
rewardlv.setInteractive()
rewardlv.init()
if (settings['debug_critic']): #True
criticLosses = []
criticRegularizationCosts = []
if (settings['visualize_learning']):
title = settings['agent_name']
k = title.rfind(".") + 1
if (k > len(title)): ## name does not contain a .
k = 0
title = title[k:]
critic_loss_viz = NNVisualize(title=str("Critic Loss") + " with " +
title)
critic_loss_viz.setInteractive()
critic_loss_viz.init()
critic_regularization_viz = NNVisualize(
title=str("Critic Reg Cost") + " with " + title)
critic_regularization_viz.setInteractive()
critic_regularization_viz.init()
if (settings['debug_actor']): # True
actorLosses = []
actorRegularizationCosts = []
if (settings['visualize_learning']): #False
title = settings['agent_name']
k = title.rfind(".") + 1
if (k > len(title)): ## name does not contain a .
k = 0
title = title[k:]
actor_loss_viz = NNVisualize(title=str("Actor Loss") + " with " +
title)
actor_loss_viz.setInteractive()
actor_loss_viz.init()
actor_regularization_viz = NNVisualize(
title=str("Actor Reg Cost") + " with " + title)
actor_regularization_viz.setInteractive()
actor_regularization_viz.init()
model = createRLAgent(settings['agent_name'], state_bounds,
discrete_actions, reward_bounds,
settings) #return a model class
forwardDynamicsModel = None
if (settings['train_forward_dynamics']): #False
if (settings['forward_dynamics_model_type'] == "SingleNet"):
print(
"Creating forward dynamics network: Using single network model"
)
forwardDynamicsModel = createForwardDynamicsModel(settings,
state_bounds,
action_bounds,
None,
None,
agentModel=model)
else:
print("Creating forward dynamics network")
forwardDynamicsModel = createForwardDynamicsModel(settings,
state_bounds,
action_bounds,
None,
None,
agentModel=None)
forwardDynamicsModel.setActor(actor)
forwardDynamicsModel.init(len(state_bounds[0]), len(action_bounds[0]),
state_bounds, action_bounds, actor, None,
settings)
(agent,
learning_workers) = createLearningAgent(settings, output_experience_queue,
state_bounds, action_bounds,
reward_bounds)
masterAgent = agent
### These are the workers for training
(sim_workers, sim_work_queues) = createSimWorkers(
settings, input_anchor_queue, output_experience_queue,
eval_episode_data_queue, model, forwardDynamicsModel, exp_val,
state_bounds, action_bounds, reward_bounds)
eval_sim_workers = sim_workers
eval_sim_work_queues = sim_work_queues
if ('override_sim_env_id' in settings
and (settings['override_sim_env_id'] != False)): #True
(eval_sim_workers, eval_sim_work_queues) = createSimWorkers(
settings,
input_anchor_queue_eval,
output_experience_queue,
eval_episode_data_queue,
model,
forwardDynamicsModel,
exp_val,
state_bounds,
action_bounds,
reward_bounds,
default_sim_id=settings['override_sim_env_id']) # id=1
else:
input_anchor_queue_eval = input_anchor_queue
best_eval = -100000000.0
best_dynamicsLosses = best_eval * -1.0
values = []
discounted_values = []
bellman_error = []
reward_over_epoc = []
dynamicsLosses = []
dynamicsRewardLosses = []
for lw in learning_workers:
print("Learning worker")
print(lw)
if (int(settings["num_available_threads"]) > 1):
for sw in sim_workers:
print("Sim worker")
print(sw)
sw.start()
if ('override_sim_env_id' in settings
and (settings['override_sim_env_id'] != False)):
for sw in eval_sim_workers:
print("Sim worker")
print(sw)
sw.start()
## This needs to be done after the simulation worker processes are created
exp_val = createEnvironment(settings["forwardDynamics_config_file"],
settings['environment_type'],
settings,
render=settings['shouldRender'],
index=0)
exp_val.setActor(actor)
exp_val.getActor().init()
exp_val.init()
### This is for a single-threaded Synchronous sim only.
if (int(settings["num_available_threads"]) == 1
): # This is okay if there is one thread only...
sim_workers[0].setEnvironment(exp_val)
sim_workers[0].start()
if ('override_sim_env_id' in settings
and (settings['override_sim_env_id'] != False)):
eval_sim_workers[0].setEnvironment(exp_val)
eval_sim_workers[0].start()
masterAgent.setPolicy(model)
if (settings['train_forward_dynamics']):
masterAgent.setForwardDynamics(forwardDynamicsModel)
tmp_p = 1.0
message = {}
if (settings['load_saved_model']):
tmp_p = settings['min_epsilon']
data = ('Update_Policy', tmp_p, model.getStateBounds(),
model.getActionBounds(), model.getRewardBounds(),
masterAgent.getPolicy().getNetworkParameters())
if (settings['train_forward_dynamics']):
data = ('Update_Policy', tmp_p, model.getStateBounds(),
model.getActionBounds(), model.getRewardBounds(),
masterAgent.getPolicy().getNetworkParameters(),
masterAgent.getForwardDynamics().getNetworkParameters())
message['type'] = 'Update_Policy'
message['data'] = data
for m_q in sim_work_queues:
print("trainModel: Sending current network parameters: ", m_q)
m_q.put(message)
if (int(settings["num_available_threads"]) == 1):
experience, state_bounds, reward_bounds, action_bounds = collectExperience(
actor,
exp_val,
model,
settings,
sim_work_queues=None,
eval_episode_data_queue=None
) #experience: state, action, nextstate, rewards,
else:
if (settings['on_policy']):
experience, state_bounds, reward_bounds, action_bounds = collectExperience(
actor,
None,
model,
settings,
sim_work_queues=sim_work_queues,
eval_episode_data_queue=eval_episode_data_queue)
else:
experience, state_bounds, reward_bounds, action_bounds = collectExperience(
actor,
None,
model,
settings,
sim_work_queues=input_anchor_queue,
eval_episode_data_queue=eval_episode_data_queue)
masterAgent.setExperience(experience)
if ('keep_seperate_fd_exp_buffer' in settings
and (settings['keep_seperate_fd_exp_buffer'])):
masterAgent.setFDExperience(copy.deepcopy(experience))
if (not validBounds(action_bounds)):
# Check that the action bounds are spcified correctly
print("Action bounds invalid: ", action_bounds)
sys.exit()
if (not validBounds(state_bounds)):
# Probably did not collect enough bootstrapping samples to get good state bounds.
print("State bounds invalid: ", state_bounds)
state_bounds = fixBounds(np.array(state_bounds))
bound_fixed = validBounds(state_bounds)
print("State bounds fixed: ", bound_fixed)
if (not validBounds(reward_bounds)):
print("Reward bounds invalid: ", reward_bounds)
sys.exit()
print("Reward History: ", experience._reward_history)
print("Action History: ", experience._action_history)
print("Action Mean: ", np.mean(experience._action_history))
print("Experience Samples: ", (experience.samples()))
if (settings["save_experience_memory"]):
print("Saving initial experience memory")
file_name = directory + getAgentName() + "_expBufferInit.hdf5"
experience.saveToFile(file_name)
if (settings['load_saved_model']
or (settings['load_saved_model']
== 'network_and_scales')): ## Transfer learning
experience.setStateBounds(copy.deepcopy(model.getStateBounds()))
experience.setRewardBounds(copy.deepcopy(model.getRewardBounds()))
experience.setActionBounds(copy.deepcopy(model.getActionBounds()))
model.setSettings(settings)
else: ## Normal
model.setStateBounds(state_bounds)
model.setActionBounds(action_bounds)
model.setRewardBounds(reward_bounds)
experience.setStateBounds(copy.deepcopy(model.getStateBounds()))
experience.setRewardBounds(copy.deepcopy(model.getRewardBounds()))
experience.setActionBounds(copy.deepcopy(model.getActionBounds()))
masterAgent_message_queue = multiprocessing.Queue(settings['epochs'])
if (settings['train_forward_dynamics']):
if (not settings['load_saved_model']):
forwardDynamicsModel.setStateBounds(state_bounds)
forwardDynamicsModel.setActionBounds(action_bounds)
forwardDynamicsModel.setRewardBounds(reward_bounds)
masterAgent.setForwardDynamics(forwardDynamicsModel)
## Now everything related to the exp memory needs to be updated
bellman_errors = []
masterAgent.setPolicy(model)
print("Master agent state bounds: ",
repr(masterAgent.getPolicy().getStateBounds()))
for sw in sim_workers: # Need to update parameter bounds for models
print("exp: ", sw._exp)
print("sw modle: ", sw._model.getPolicy())
## If not on policy
if (not settings['on_policy']):
for lw in learning_workers:
lw._agent.setPolicy(model)
lw.setMasterAgentMessageQueue(masterAgent_message_queue)
lw.updateExperience(experience)
print("ls policy: ", lw._agent.getPolicy())
lw.start()
tmp_p = 1.0
if (settings['load_saved_model']):
tmp_p = settings['min_epsilon']
data = ('Update_Policy', tmp_p, model.getStateBounds(),
model.getActionBounds(), model.getRewardBounds(),
masterAgent.getPolicy().getNetworkParameters())
if (settings['train_forward_dynamics']):
data = ('Update_Policy', tmp_p, model.getStateBounds(),
model.getActionBounds(), model.getRewardBounds(),
masterAgent.getPolicy().getNetworkParameters(),
masterAgent.getForwardDynamics().getNetworkParameters())
message['type'] = 'Update_Policy'
message['data'] = data
for m_q in sim_work_queues:
print("trainModel: Sending current network parameters: ", m_q)
m_q.put(message)
del model
## Give gloabl access to processes to they can be terminated when ctrl+c is pressed
global sim_processes
sim_processes = sim_workers
global learning_processes
learning_processes = learning_workers
global _input_anchor_queue
_input_anchor_queue = input_anchor_queue
global _output_experience_queue
_output_experience_queue = output_experience_queue
global _eval_episode_data_queue
_eval_episode_data_queue = eval_episode_data_queue
global _sim_work_queues
_sim_work_queues = sim_work_queues
trainData = {}
trainData["mean_reward"] = []
trainData["std_reward"] = []
trainData["mean_bellman_error"] = []
trainData["std_bellman_error"] = []
trainData["mean_discount_error"] = []
trainData["std_discount_error"] = []
trainData["mean_forward_dynamics_loss"] = []
trainData["std_forward_dynamics_loss"] = []
trainData["mean_forward_dynamics_reward_loss"] = []
trainData["std_forward_dynamics_reward_loss"] = []
trainData["mean_eval"] = []
trainData["std_eval"] = []
trainData["mean_critic_loss"] = []
trainData["std_critic_loss"] = []
trainData["mean_critic_regularization_cost"] = []
trainData["std_critic_regularization_cost"] = []
trainData["mean_actor_loss"] = []
trainData["std_actor_loss"] = []
trainData["mean_actor_regularization_cost"] = []
trainData["std_actor_regularization_cost"] = []
trainData["anneal_p"] = []
if (False):
print("State Bounds:", masterAgent.getStateBounds())
print("Action Bounds:", masterAgent.getActionBounds())
print("Exp State Bounds: ", experience.getStateBounds())
print("Exp Action Bounds: ", experience.getActionBounds())
print("Starting first round")
if (settings['on_policy']):
sim_epochs_ = epochs
for round_ in range(
0, rounds): #annel value # the parameter of greedy exploration
if ('annealing_schedule' in settings
and (settings['annealing_schedule'] != False)):
p = anneal_value(float(round_ / rounds), settings_=settings)
else:
p = ((settings['initial_temperature'] / math.log(round_ + 2)))
p = max(settings['min_epsilon'],
min(settings['epsilon'], p)) # Keeps it between 1.0 and 0.2
if (settings['load_saved_model']):
p = settings['min_epsilon']
for epoch in range(epochs):
if (settings['on_policy']):
out = simModelParrallel(
sw_message_queues=sim_work_queues,
model=masterAgent,
settings=settings,
eval_episode_data_queue=eval_episode_data_queue,
anchors=settings['num_on_policy_rollouts'])
(
tuples, discounted_sum, q_value, evalData
) = out # tuples = states, actions, result_states, rewards, falls, G_ts, advantage, exp_actions
(__states, __actions, __result_states, __rewards, __falls,
__G_ts, advantage__, exp_actions__) = tuples
for i in range(1):
masterAgent.train(_states=__states,
_actions=__actions,
_rewards=__rewards,
_result_states=__result_states,
_falls=__falls,
_advantage=advantage__,
_exp_actions=exp_actions__)
if (('anneal_on_policy' in settings)
and settings['anneal_on_policy']):
p_tmp_ = p
else:
p_tmp_ = 1.0
data = ('Update_Policy', p_tmp_, masterAgent.getStateBounds(),
masterAgent.getActionBounds(),
masterAgent.getRewardBounds(),
masterAgent.getPolicy().getNetworkParameters())
message = {}
message['type'] = 'Update_Policy'
message['data'] = data
if (settings['train_forward_dynamics']):
data = ('Update_Policy', p_tmp_,
masterAgent.getStateBounds(),
masterAgent.getActionBounds(),
masterAgent.getRewardBounds(),
masterAgent.getPolicy().getNetworkParameters(),
masterAgent.getForwardDynamics().
getNetworkParameters())
message['data'] = data
for m_q in sim_work_queues:
## block on full queue
m_q.put(message)
if ('override_sim_env_id' in settings
and (settings['override_sim_env_id'] != False)):
for m_q in eval_sim_work_queues:
## block on full queue
m_q.put(message)
else:
episodeData = {}
episodeData['data'] = epoch
episodeData['type'] = 'sim'
input_anchor_queue.put(episodeData)
if masterAgent.getExperience().samples(
) >= batch_size: #更新policy网络
states, actions, result_states, rewards, falls, G_ts, exp_actions = masterAgent.getExperience(
).get_batch(batch_size)
error = masterAgent.bellman_error(states, actions, rewards,
result_states, falls)
bellman_errors.append(error)
if (settings['debug_critic']):
loss__ = masterAgent.getPolicy()._get_critic_loss(
) # uses previous call batch data
criticLosses.append(loss__)
regularizationCost__ = masterAgent.getPolicy(
)._get_critic_regularization()
criticRegularizationCosts.append(regularizationCost__)
if (settings['debug_actor']): #True
loss__ = masterAgent.getPolicy()._get_actor_loss(
) # uses previous call batch data
actorLosses.append(loss__)
regularizationCost__ = masterAgent.getPolicy(
)._get_actor_regularization()
actorRegularizationCosts.append(regularizationCost__)
if not all(np.isfinite(error)):
print(
"States: " + str(states) + " ResultsStates: " +
str(result_states) + " Rewards: " + str(rewards) +
" Actions: " + str(actions) + " Falls: ", str(falls))
print("Bellman Error is Nan: " + str(error) +
str(np.isfinite(error)))
sys.exit()
error = np.mean(np.fabs(error))
if error > 10000:
print("Error to big: ")
print(states, actions, rewards, result_states)
if (settings['train_forward_dynamics']): #False
dynamicsLoss = masterAgent.getForwardDynamics(
).bellman_error(states, actions, result_states, rewards)
dynamicsLoss = np.mean(np.fabs(dynamicsLoss)) #fabs：计算绝对值
dynamicsLosses.append(dynamicsLoss)
if (settings['train_reward_predictor']):
dynamicsRewardLoss = masterAgent.getForwardDynamics(
).reward_error(states, actions, result_states, rewards)
dynamicsRewardLoss = np.mean(
np.fabs(dynamicsRewardLoss))
dynamicsRewardLosses.append(dynamicsRewardLoss)
if (settings['train_forward_dynamics']):
print("Round: " + str(round_) + " Epoch: " + str(epoch) +
" p: " + str(p) + " With mean reward: " +
str(np.mean(rewards)) + " bellman error: " +
str(error) + " ForwardPredictionLoss: " +
str(dynamicsLoss))
else:
print("Round: " + str(round_) + " Epoch: " + str(epoch) +
" p: " + str(p) + " With mean reward: " +
str(np.mean(rewards)) + " bellman error: " +
str(error))
if (settings["print_levels"][settings["print_level"]] >=
settings["print_levels"]['train']):
print("Master agent experience size: " +
str(masterAgent.getExperience().samples()))
if (not settings['on_policy']):
## There could be stale policy parameters in here, use the last set put in the queue
data = None
while (not masterAgent_message_queue.empty()):
## Don't block
try:
data = masterAgent_message_queue.get(False)
except Exception as inst:
pass
if (not (data == None)):
masterAgent.setExperience(data[0])
masterAgent.getPolicy().setNetworkParameters(data[1])
masterAgent.setStateBounds(
masterAgent.getExperience().getStateBounds())
masterAgent.setActionBounds(
masterAgent.getExperience().getActionBounds())
masterAgent.setRewardBounds(
masterAgent.getExperience().getRewardBounds())
if (settings['train_forward_dynamics']):
masterAgent.getForwardDynamics().setNetworkParameters(
data[2])
if ('keep_seperate_fd_exp_buffer' in settings
and (settings['keep_seperate_fd_exp_buffer'])):
masterAgent.setFDExperience(data[3])
# this->_actor->iterate();
## This will let me know which part of learning is going slower training updates or simulation
if (settings["print_levels"][settings["print_level"]] >=
settings["print_levels"]['train']):
print("sim queue size: ", input_anchor_queue.qsize()) #返回队列的大小
if (output_experience_queue != None):
print("exp tuple queue size: ", output_experience_queue.qsize())
if (not settings['on_policy']):
data = ('Update_Policy', p, masterAgent.getStateBounds(),
masterAgent.getActionBounds(),
masterAgent.getRewardBounds(),
masterAgent.getPolicy().getNetworkParameters())
if (settings['train_forward_dynamics']):
data = (
'Update_Policy', p, masterAgent.getStateBounds(),
masterAgent.getActionBounds(),
masterAgent.getRewardBounds(),
masterAgent.getPolicy().getNetworkParameters(),
masterAgent.getForwardDynamics().getNetworkParameters())
message['type'] = 'Update_Policy'
message['data'] = data
for m_q in sim_work_queues:
## Don't block on full queue
try:
m_q.put(message, False)
except:
print("SimWorker model parameter message queue full: ",
m_q.qsize())
if ('override_sim_env_id' in settings
and (settings['override_sim_env_id'] != False)):
for m_q in eval_sim_work_queues:
## Don't block on full queue
try:
m_q.put(message, False)
except:
print("SimWorker model parameter message queue full: ",
m_q.qsize())
if (round_ % settings['plotting_update_freq_num_rounds']) == 0:
# Running less often helps speed learning up.
# Sync up sim actors
if (settings['on_policy']):
mean_reward, std_reward, mean_bellman_error, std_bellman_error, mean_discount_error, std_discount_error, mean_eval, std_eval = evalModelParrallel(
input_anchor_queue=eval_sim_work_queues,
model=masterAgent,
settings=settings,
eval_episode_data_queue=eval_episode_data_queue,
anchors=settings['eval_epochs'])
else:
mean_reward, std_reward, mean_bellman_error, std_bellman_error, mean_discount_error, std_discount_error, mean_eval, std_eval = evalModelParrallel(
input_anchor_queue=input_anchor_queue_eval,
model=masterAgent,
settings=settings,
eval_episode_data_queue=eval_episode_data_queue,
anchors=settings['eval_epochs'])
print(mean_reward, std_reward, mean_bellman_error,
std_bellman_error, mean_discount_error, std_discount_error)
if mean_bellman_error > 10000:
print("Error to big: ")
else:
if (settings['train_forward_dynamics']): #false
mean_dynamicsLosses = np.mean(dynamicsLosses)
std_dynamicsLosses = np.std(dynamicsLosses)
dynamicsLosses = []
if (settings['train_reward_predictor']): #false
mean_dynamicsRewardLosses = np.mean(dynamicsRewardLosses)
std_dynamicsRewardLosses = np.std(dynamicsRewardLosses)
dynamicsRewardLosses = []
trainData["mean_reward"].append(mean_reward)
trainData["std_reward"].append(std_reward)
trainData["anneal_p"].append(p)
trainData["mean_bellman_error"].append(
np.mean(np.fabs(bellman_errors)))
trainData["std_bellman_error"].append(np.std(bellman_errors))
bellman_errors = []
trainData["mean_discount_error"].append(mean_discount_error)
trainData["std_discount_error"].append(std_discount_error)
trainData["mean_eval"].append(mean_eval)
trainData["std_eval"].append(std_eval)
if (settings['train_forward_dynamics']):
trainData["mean_forward_dynamics_loss"].append(
mean_dynamicsLosses)
trainData["std_forward_dynamics_loss"].append(
std_dynamicsLosses)
if (settings['train_reward_predictor']):
trainData["mean_forward_dynamics_reward_loss"].append(
mean_dynamicsRewardLosses)
trainData["std_forward_dynamics_reward_loss"].append(
std_dynamicsRewardLosses)
if (round_ % settings['saving_update_freq_num_rounds']) == 0:
if (settings['train_forward_dynamics']):
file_name_dynamics = directory + "forward_dynamics_" + ".pkl"
f = open(file_name_dynamics, 'wb')
dill.dump(masterAgent.getForwardDynamics(), f)
f.close()
if mean_dynamicsLosses < best_dynamicsLosses:
best_dynamicsLosses = mean_dynamicsLosses
print("Saving BEST current forward dynamics agent: " +
str(best_dynamicsLosses))
file_name_dynamics = directory + "forward_dynamics_" + "_Best.pkl"
f = open(file_name_dynamics, 'wb')
dill.dump(masterAgent.getForwardDynamics(), f) #save model
f.close()
if (mean_eval > best_eval):
best_eval = mean_eval
print("Saving BEST current agent: " + str(best_eval))
file_name = directory + getAgentName() + "_Best.pkl"
f = open(file_name, 'wb')
dill.dump(masterAgent.getPolicy(), f)
f.close()
if settings['save_trainData']:
fp = open(
directory + "trainingData_" + str(settings['agent_name']) +
".json", 'w')
## because json does not serialize np.float32
for key in trainData:
trainData[key] = [float(i) for i in trainData[key]]
json.dump(trainData, fp)
fp.close()
print("Saving current masterAgent")
file_name = directory + getAgentName() + ".pkl"
f = open(file_name, 'wb')
dill.dump(masterAgent.getPolicy(), f)
f.close()
gc.collect()
print("Terminating Workers")
if (settings['on_policy']):
for m_q in sim_work_queues:
## block on full queue
m_q.put(None)
if ('override_sim_env_id' in settings
and (settings['override_sim_env_id'] != False)):
for m_q in eval_sim_work_queues:
## block on full queue
m_q.put(None)
for sw in sim_workers: # Should update these more often
sw.join()
if ('override_sim_env_id' in settings
and (settings['override_sim_env_id'] != False)):
for sw in eval_sim_workers: # Should update these more often
sw.join()
for i in range(len(sim_work_queues)):
print("sim_work_queues size: ", sim_work_queues[i].qsize())
while (not sim_work_queues[i].empty()): ### Empty the queue
## Don't block
try:
data_ = sim_work_queues[i].get(False)
except Exception as inst:
pass
print("sim_work_queues size: ", sim_work_queues[i].qsize())
for i in range(len(eval_sim_work_queues)):
print("eval_sim_work_queues size: ", eval_sim_work_queues[i].qsize())
while (not eval_sim_work_queues[i].empty()): ### Empty the queue
## Don't block
try:
data_ = eval_sim_work_queues[i].get(False)
except Exception as inst:
pass
print("eval_sim_work_queues size: ", eval_sim_work_queues[i].qsize())
print("Finish sim")
exp_val.finish()
print("Save last versions of files.")
file_name = directory + getAgentName() + ".pkl"
f = open(file_name, 'wb')
dill.dump(masterAgent.getPolicy(), f)
f.close()
f = open(
directory + "trainingData_" + str(settings['agent_name']) + ".json",
"w")
for key in trainData:
trainData[key] = [float(i) for i in trainData[key]]
json.dump(trainData, f, sort_keys=True, indent=4)
f.close()
if (settings['train_forward_dynamics']):
file_name_dynamics = directory + "forward_dynamics_" + ".pkl"
f = open(file_name_dynamics, 'wb')
dill.dump(masterAgent.getForwardDynamics(), f)
f.close()
print("Delete any plots being used")
gc.collect() #立即释放内存
def fitModelToData(settingsFileName):
"""
State is the input state and Action is the desired output (y).
"""
# from model.ModelUtil import *
file = open(settingsFileName)
settings = json.load(file)
print("Settings: " + str(json.dumps(settings)))
file.close()
import os
os.environ['THEANO_FLAGS'] = "mode=FAST_RUN,device=" + settings[
'training_processor_type'] + ",floatX=" + settings['float_type']
## Theano needs to be imported after the flags are set.
# from ModelEvaluation import *
# from model.ModelUtil import *
# print ( "theano.config.mode: ", theano.config.mode)
from ModelEvaluation import SimWorker, evalModelParrallel, collectExperience, simEpoch, evalModel
from model.ModelUtil import validBounds
from model.LearningAgent import LearningAgent, LearningWorker
from util.SimulationUtil import validateSettings, createEnvironment, createRLAgent, createActor
from util.SimulationUtil import getDataDirectory, createForwardDynamicsModel, createSampler
from util.ExperienceMemory import ExperienceMemory
from RLVisualize import RLVisualize
from NNVisualize import NNVisualize
from sim.PendulumEnvState import PendulumEnvState
from sim.PendulumEnv import PendulumEnv
from sim.BallGame2DEnv import BallGame2DEnv
import time
settings = validateSettings(settings)
# anchor_data_file = open(settings["anchor_file"])
# _anchors = getAnchors(anchor_data_file)
# print ("Length of anchors epochs: ", str(len(_anchors)))
# anchor_data_file.close()
train_forward_dynamics = True
model_type = settings["model_type"]
directory = getDataDirectory(settings)
discrete_actions = np.array(settings['discrete_actions'])
num_actions = discrete_actions.shape[0] # number of rows
rounds = settings["rounds"]
epochs = settings["epochs"]
# num_states=settings["num_states"]
epsilon = settings["epsilon"]
discount_factor = settings["discount_factor"]
# max_reward=settings["max_reward"]
reward_bounds = np.array(settings["reward_bounds"])
batch_size = settings["batch_size"]
train_on_validation_set = settings["train_on_validation_set"]
state_bounds = np.array(settings['state_bounds'])
discrete_actions = np.array(settings['discrete_actions'])
print("Sim config file name: ", str(settings["sim_config_file"]))
# c = characterSim.Configuration(str(settings["sim_config_file"]))
# c = characterSim.Configuration("../data/epsilon0Config.ini")
action_space_continuous = settings['action_space_continuous']
# states2 = np.transpose(np.repeat([states], 2, axis=0))
# print states2
if action_space_continuous:
action_bounds = np.array(settings["action_bounds"], dtype=float)
if action_space_continuous:
experience = ExperienceMemory(len(state_bounds[0]),
len(action_bounds[0]),
settings['expereince_length'],
continuous_actions=True,
settings=settings)
else:
experience = ExperienceMemory(len(state_bounds[0]), 1,
settings['expereince_length'])
file_name = directory + getAgentName() + "expBufferInit.hdf5"
# experience.saveToFile(file_name)
experience.loadFromFile(file_name)
state_bounds = experience._state_bounds
action_bounds = experience._action_bounds
reward_bounds = experience._reward_bounds
output_experience_queue = multiprocessing.Queue(
settings['queue_size_limit'])
mgr = multiprocessing.Manager()
namespace = mgr.Namespace()
learning_workers = []
# for process in range(settings['num_available_threads']):
for process in range(1):
# this is the process that selects which game to play
agent = LearningAgent(n_in=len(state_bounds[0]),
n_out=len(action_bounds[0]),
state_bounds=state_bounds,
action_bounds=action_bounds,
reward_bound=reward_bounds,
settings_=settings)
agent.setSettings(settings)
"""
if action_space_continuous:
model = createRLAgent(settings['agent_name'], state_bounds, action_bounds, reward_bounds, settings)
else:
model = createRLAgent(settings['agent_name'], state_bounds, discrete_actions, reward_bounds, settings)
model.setStateBounds(state_bounds)
model.setActionBounds(action_bounds)
model.setRewardBounds(reward_bounds)
"""
# agent.setPolicy(model)
# actor.setPolicy(model)
# agent.setExperience(experience)
# namespace.agentPoly = agent.getPolicy().getNetworkParameters()
# namespace.experience = experience
lw = LearningWorker(output_experience_queue, agent, namespace)
# lw.start()
learning_workers.append(lw)
masterAgent = agent
masterAgent.setExperience(experience)
if action_space_continuous:
model = createRLAgent(settings['agent_name'], state_bounds,
action_bounds, reward_bounds, settings)
else:
model = createRLAgent(settings['agent_name'], state_bounds,
discrete_actions, reward_bounds, settings)
if (not settings['load_saved_model']):
model.setStateBounds(state_bounds)
model.setActionBounds(action_bounds)
model.setRewardBounds(reward_bounds)
else: # continuation learning
experience.setStateBounds(model.getStateBounds())
experience.setRewardBounds(model.getRewardBounds())
experience.setActionBounds(model.getActionBounds())
if (settings['train_forward_dynamics']):
print("Created forward dynamics network")
# forwardDynamicsModel = ForwardDynamicsNetwork(state_length=len(state_bounds[0]),action_length=len(action_bounds[0]), state_bounds=state_bounds, action_bounds=action_bounds, settings_=settings)
forwardDynamicsModel = createForwardDynamicsModel(
settings, state_bounds, action_bounds, None, None)
masterAgent.setForwardDynamics(forwardDynamicsModel)
forwardDynamicsModel.setActor(actor)
# forwardDynamicsModel.setEnvironment(exp)
forwardDynamicsModel.init(len(state_bounds[0]), len(action_bounds[0]),
state_bounds, action_bounds, actor, None,
settings)
namespace.forwardNN = masterAgent.getForwardDynamics(
).getNetworkParameters()
# actor.setForwardDynamicsModel(forwardDynamicsModel)
namespace.forwardDynamicsModel = forwardDynamicsModel
## Now everything related to the exp memory needs to be updated
bellman_errors = []
masterAgent.setPolicy(model)
# masterAgent.setForwardDynamics(forwardDynamicsModel)
namespace.agentPoly = masterAgent.getPolicy().getNetworkParameters()
namespace.model = model
# experience = ExperienceMemory(len(state_bounds[0]), len(action_bounds[0]), experience_length, continuous_actions=True)
"""
for i in range(experience_length):
action_ = np.array([actions[i]])
state_ = np.array([states[i]])
# print "Action: " + str([actions[i]])
experience.insert(norm_state(state_, state_bounds), norm_action(action_, action_bounds),
norm_state(state_, state_bounds), norm_reward(np.array([0]), reward_bounds))
"""
if (settings['visualize_learning']):
rlv = NNVisualize(title=str(directory), settings=settings)
rlv.setInteractive()
rlv.init()
if (settings['debug_critic']):
criticLosses = []
criticRegularizationCosts = []
if (settings['visualize_learning']):
critic_loss_viz = NNVisualize(title=str("Critic Loss") + " with " +
str(settings["model_type"]))
critic_loss_viz.setInteractive()
critic_loss_viz.init()
critic_regularization_viz = NNVisualize(
title=str("Critic Regularization Cost") + " with " +
str(settings["model_type"]))
critic_regularization_viz.setInteractive()
critic_regularization_viz.init()
if (settings['debug_actor']):
actorLosses = []
actorRegularizationCosts = []
if (settings['visualize_learning']):
actor_loss_viz = NNVisualize(title=str("Actor Loss") + " with " +
str(settings["model_type"]))
actor_loss_viz.setInteractive()
actor_loss_viz.init()
actor_regularization_viz = NNVisualize(
title=str("Actor Regularization Cost") + " with " +
str(settings["model_type"]))
actor_regularization_viz.setInteractive()
actor_regularization_viz.init()
trainData = {}
trainData["mean_reward"] = []
trainData["std_reward"] = []
trainData["mean_bellman_error"] = []
trainData["std_bellman_error"] = []
trainData["mean_discount_error"] = []
trainData["std_discount_error"] = []
trainData["mean_forward_dynamics_loss"] = []
trainData["std_forward_dynamics_loss"] = []
trainData["mean_eval"] = []
trainData["std_eval"] = []
trainData["mean_critic_loss"] = []
trainData["std_critic_loss"] = []
trainData["mean_critic_regularization_cost"] = []
trainData["std_critic_regularization_cost"] = []
trainData["mean_actor_loss"] = []
trainData["std_actor_loss"] = []
trainData["mean_actor_regularization_cost"] = []
trainData["std_actor_regularization_cost"] = []
# dynamicsLosses=[]
best_dynamicsLosses = 1000000
_states, _actions, _result_states, _rewards, _falls, _G_ts = experience.get_batch(
batch_size)
"""
_states = theano.shared(np.array(_states, dtype=theano.config.floatX))
_actions = theano.shared(np.array(_actions, dtype=theano.config.floatX))
_result_states = theano.shared(np.array(_result_states, dtype=theano.config.floatX))
_rewards = theano.shared(np.array(_rewards, dtype=theano.config.floatX))
"""
for round_ in range(rounds):
t0 = time.time()
# out = simEpoch(actor, exp_val, masterAgent, discount_factor, anchors=epoch, action_space_continuous=action_space_continuous, settings=settings,
# print_data=False, p=1.0, validation=False, epoch=epoch, evaluation=False, _output_queue=None )
# (tuples, discounted_sum, q_value, evalData) = out
# (__states, __actions, __result_states, __rewards, __falls, __G_ts) = tuples
__states, __actions, __result_states, __rewards, __falls, __G_ts = experience.get_batch(
100)
# print("**** training states: ", np.array(__states).shape)
# print("**** training __result_states: ", np.array(__result_states).shape)
# print ("Actions before: ", __actions)
for i in range(1):
masterAgent.train(_states=__states,
_actions=__actions,
_rewards=__rewards,
_result_states=__result_states,
_falls=__falls)
t1 = time.time()
time_taken = t1 - t0
if masterAgent.getExperience().samples() > batch_size:
states, actions, result_states, rewards, falls, G_ts = masterAgent.getExperience(
).get_batch(batch_size)
print("Batch size: " + str(batch_size))
error = masterAgent.bellman_error(states, actions, rewards,
result_states, falls)
bellman_errors.append(error)
if (settings['debug_critic']):
loss__ = masterAgent.getPolicy()._get_critic_loss(
) # uses previous call batch data
criticLosses.append(loss__)
regularizationCost__ = masterAgent.getPolicy(
)._get_critic_regularization()
criticRegularizationCosts.append(regularizationCost__)
if (settings['debug_actor']):
"""
print( "Advantage: ", masterAgent.getPolicy()._get_advantage())
print("Policy prob: ", masterAgent.getPolicy()._q_action())
print("Policy log prob: ", masterAgent.getPolicy()._get_log_prob())
print( "Actor loss: ", masterAgent.getPolicy()._get_action_diff())
"""
loss__ = masterAgent.getPolicy()._get_actor_loss(
) # uses previous call batch data
actorLosses.append(loss__)
regularizationCost__ = masterAgent.getPolicy(
)._get_actor_regularization()
actorRegularizationCosts.append(regularizationCost__)
if not all(np.isfinite(error)):
print(
"States: " + str(states) + " ResultsStates: " +
str(result_states) + " Rewards: " + str(rewards) +
" Actions: " + str(actions) + " Falls: ", str(falls))
print("Bellman Error is Nan: " + str(error) +
str(np.isfinite(error)))
sys.exit()
error = np.mean(np.fabs(error))
if error > 10000:
print("Error to big: ")
print(states, actions, rewards, result_states)
if (settings['train_forward_dynamics']):
dynamicsLoss = masterAgent.getForwardDynamics().bellman_error(
states, actions, result_states, rewards)
dynamicsLoss = np.mean(np.fabs(dynamicsLoss))
dynamicsLosses.append(dynamicsLoss)
if (settings['train_forward_dynamics']):
print("Round: " + str(round_) + " bellman error: " +
str(error) + " ForwardPredictionLoss: " +
str(dynamicsLoss) + " in " + str(time_taken) +
" seconds")
else:
print("Round: " + str(round_) + " bellman error: " +
str(error) + " in " + str(time_taken) + " seconds")
# discounted_values.append(discounted_sum)
print("Master agent experience size: " +
str(masterAgent.getExperience().samples()))
# print ("**** Master agent experience size: " + str(learning_workers[0]._agent._expBuff.samples()))
# masterAgent.getPolicy().setNetworkParameters(namespace.agentPoly)
# masterAgent.setExperience(learningNamespace.experience)
# if (settings['train_forward_dynamics']):
# masterAgent.getForwardDynamics().setNetworkParameters(namespace.forwardNN)
"""
for sw in sim_workers: # Should update these more often?
sw._model.getPolicy().setNetworkParameters(namespace.agentPoly)
if (settings['train_forward_dynamics']):
sw._model.getForwardDynamics().setNetworkParameters(namespace.forwardNN)
"""
# experience = learningNamespace.experience
# actor.setExperience(experience)
"""
pr.disable()
f = open('x.prof', 'a')
pstats.Stats(pr, stream=f).sort_stats('time').print_stats()
f.close()
"""
trainData["mean_bellman_error"].append(np.mean(
np.fabs(bellman_errors)))
trainData["std_bellman_error"].append(np.std(bellman_errors))
if (settings['visualize_learning']):
rlv.updateLoss(np.array(trainData["mean_bellman_error"]),
np.array(trainData["std_bellman_error"]))
rlv.redraw()
rlv.setInteractiveOff()
rlv.saveVisual(directory + "trainingGraphNN")
rlv.setInteractive()
# print "Error: " + str(error)
if (settings['debug_critic']):
mean_criticLosses = np.mean(criticLosses)
std_criticLosses = np.std(criticLosses)
trainData["mean_critic_loss"].append(mean_criticLosses)
trainData["std_critic_loss"].append(std_criticLosses)
criticLosses = []
if (settings['visualize_learning']):
critic_loss_viz.updateLoss(
np.array(trainData["mean_critic_loss"]),
np.array(trainData["std_critic_loss"]))
critic_loss_viz.redraw()
critic_loss_viz.setInteractiveOff()
critic_loss_viz.saveVisual(directory + "criticLossGraph")
critic_loss_viz.setInteractive()
mean_criticRegularizationCosts = np.mean(criticRegularizationCosts)
std_criticRegularizationCosts = np.std(criticRegularizationCosts)
trainData["mean_critic_regularization_cost"].append(
mean_criticRegularizationCosts)
trainData["std_critic_regularization_cost"].append(
std_criticRegularizationCosts)
criticRegularizationCosts = []
if (settings['visualize_learning']):
critic_regularization_viz.updateLoss(
np.array(trainData["mean_critic_regularization_cost"]),
np.array(trainData["std_critic_regularization_cost"]))
critic_regularization_viz.redraw()
critic_regularization_viz.setInteractiveOff()
critic_regularization_viz.saveVisual(
directory + "criticRegularizationGraph")
critic_regularization_viz.setInteractive()
if (settings['debug_actor']):
mean_actorLosses = np.mean(actorLosses)
std_actorLosses = np.std(actorLosses)
trainData["mean_actor_loss"].append(mean_actorLosses)
trainData["std_actor_loss"].append(std_actorLosses)
actorLosses = []
if (settings['visualize_learning']):
actor_loss_viz.updateLoss(
np.array(trainData["mean_actor_loss"]),
np.array(trainData["std_actor_loss"]))
actor_loss_viz.redraw()
actor_loss_viz.setInteractiveOff()
actor_loss_viz.saveVisual(directory + "actorLossGraph")
actor_loss_viz.setInteractive()
mean_actorRegularizationCosts = np.mean(actorRegularizationCosts)
std_actorRegularizationCosts = np.std(actorRegularizationCosts)
trainData["mean_actor_regularization_cost"].append(
mean_actorRegularizationCosts)
trainData["std_actor_regularization_cost"].append(
std_actorRegularizationCosts)
actorRegularizationCosts = []
if (settings['visualize_learning']):
actor_regularization_viz.updateLoss(
np.array(trainData["mean_actor_regularization_cost"]),
np.array(trainData["std_actor_regularization_cost"]))
actor_regularization_viz.redraw()
actor_regularization_viz.setInteractiveOff()
actor_regularization_viz.saveVisual(directory +
"actorRegularizationGraph")
actor_regularization_viz.setInteractive()
class QProp(AlgorithmInterface):
def __init__(self, model, n_in, n_out, state_bounds, action_bounds,
reward_bound, settings_):
"""
In order to get this to work we need to be careful not to update the actor parameters
when updating the critic. This can be an issue when the Concatenating networks together.
The first first network becomes a part of the second. However you can still access the first
network by itself but an updates on the second network will effect the first network.
Care needs to be taken to make sure only the parameters of the second network are updated.
"""
super(QProp, self).__init__(model, n_in, n_out, state_bounds,
action_bounds, reward_bound, settings_)
# if ('train_extra_value_function' in self.getSettings() and (self.getSettings()['train_extra_value_function'] == True)):
self._experience = ExperienceMemory(
n_in,
n_out,
self.getSettings()['expereince_length'],
continuous_actions=True,
settings=self.getSettings())
self._experience.setStateBounds(copy.deepcopy(self.getStateBounds()))
self._experience.setRewardBounds(copy.deepcopy(self.getRewardBounds()))
self._experience.setActionBounds(copy.deepcopy(self.getActionBounds()))
self._use_basic_polcy_grad = False
self._Fallen = T.bcol("Fallen")
## because float64 <= float32 * int32, need to use int16 or int8
self._Fallen.tag.test_value = np.zeros((self._batch_size, 1),
dtype=np.dtype('int8'))
self._fallen_shared = theano.shared(np.zeros((self._batch_size, 1),
dtype='int8'),
broadcastable=(False, True))
self._Action = T.matrix("Action2")
self._Action.tag.test_value = np.random.rand(self._batch_size,
self._action_length)
self._Tmp_Target = T.col("Tmp_Target")
self._Tmp_Target.tag.test_value = np.zeros(
(self._batch_size, 1),
dtype=np.dtype(self.getSettings()['float_type']))
self._tmp_target_shared = theano.shared(np.zeros(
(self._batch_size, 1), dtype=self.getSettings()['float_type']),
broadcastable=(False, True))
self._Advantage = T.col("Advantage")
self._Advantage.tag.test_value = np.zeros(
(self._batch_size, 1),
dtype=np.dtype(self.getSettings()['float_type']))
self._advantage_shared = theano.shared(np.zeros(
(self._batch_size, 1), dtype=self.getSettings()['float_type']),
broadcastable=(False, True))
self._QProp_N = T.col("QProp_N")
self._QProp_N.tag.test_value = np.zeros(
(self._batch_size, 1),
dtype=np.dtype(self.getSettings()['float_type']))
self._QProp_N_shared = theano.shared(np.zeros(
(self._batch_size, 1), dtype=self.getSettings()['float_type']),
broadcastable=(False, True))
self._modelTarget = copy.deepcopy(model)
self._modelTarget2 = copy.deepcopy(model)
self._learning_rate = self.getSettings()['learning_rate']
self._discount_factor = self.getSettings()['discount_factor']
self._rho = self.getSettings()['rho']
self._rms_epsilon = self.getSettings()['rms_epsilon']
self._weight_update_steps = self.getSettings(
)['steps_until_target_network_update']
self._updates = 0
self._decay_weight = self.getSettings()['regularization_weight']
self._critic_regularization_weight = self.getSettings(
)["critic_regularization_weight"]
self._critic_learning_rate = self.getSettings()["critic_learning_rate"]
self._q_valsActA = lasagne.layers.get_output(
self._model.getActorNetwork(),
self._model.getStateSymbolicVariable(),
deterministic=True)
self._q_valsActTarget = lasagne.layers.get_output(
self._modelTarget.getActorNetwork(),
self._model.getResultStateSymbolicVariable(),
deterministic=True)
self._q_valsActTarget_State = lasagne.layers.get_output(
self._modelTarget2.getActorNetwork(),
self._model.getStateSymbolicVariable(),
deterministic=True)
self._q_valsActASTD = (T.ones_like(
self._q_valsActA)) * self.getSettings()['exploration_rate']
self._q_valsActTargetSTD = (T.ones_like(self._q_valsActTarget_State)
) * self.getSettings()['exploration_rate']
inputs_1 = {
self._model.getStateSymbolicVariable(): self._model.getStates(),
self._model.getActionSymbolicVariable(): self._model.getActions()
}
self._q_valsA = lasagne.layers.get_output(
self._model.getCriticNetwork(), inputs_1)
inputs_2 = {
self._modelTarget.getStateSymbolicVariable():
self._model.getResultStates(),
self._modelTarget.getActionSymbolicVariable():
self._model.getActions()
}
self._q_valsB_ = lasagne.layers.get_output(
self._modelTarget.getCriticNetwork(), inputs_2, deterministic=True)
self._q_func = self._q_valsA
self._q_funcB = self._q_valsB_
self._q_funcAct = self._q_valsActA
self._diff = self._Tmp_Target - self._q_func
loss = T.pow(self._diff, 2)
self._loss = T.mean(loss)
self._params = lasagne.layers.helper.get_all_params(
self._model.getCriticNetwork())
print("******Number of Layers is: " + str(
len(
lasagne.layers.helper.get_all_params(
self._model.getCriticNetwork()))))
print("******Number of Action Layers is: " + str(
len(
lasagne.layers.helper.get_all_params(
self._model.getActorNetwork()))))
self._actionParams = lasagne.layers.helper.get_all_params(
self._model.getActorNetwork())
self._givens_ = {
self._model.getStateSymbolicVariable(): self._model.getStates(),
self._model.getActionSymbolicVariable(): self._model.getActions(),
self._Tmp_Target: self._tmp_target_shared
}
self._actGivens = {
self._model.getStateSymbolicVariable(): self._model.getStates(),
}
self._critic_regularization = (
self._critic_regularization_weight *
lasagne.regularization.regularize_network_params(
self._model.getCriticNetwork(), lasagne.regularization.l2))
## MSE update
self._value_grad = T.grad(self._loss + self._critic_regularization,
self._params)
print("Optimizing Value Function with ",
self.getSettings()['optimizer'], " method")
self._updates_ = lasagne.updates.adam(self._value_grad,
self._params,
self._critic_learning_rate,
beta1=0.9,
beta2=0.9,
epsilon=self._rms_epsilon)
self._givens_grad = {
self._model.getStateSymbolicVariable(): self._model.getStates(),
self._model.getActionSymbolicVariable(): self._model.getActions(),
}
## Some cool stuff to backprop action gradients
self._action_grad = T.matrix("Action_Grad")
self._action_grad.tag.test_value = np.zeros(
(self._batch_size, self._action_length),
dtype=np.dtype(self.getSettings()['float_type']))
self._action_grad_shared = theano.shared(
np.zeros((self._batch_size, self._action_length),
dtype=self.getSettings()['float_type']))
### Maximize wrt q function
self._action_mean_grads = T.grad(
cost=None,
wrt=self._actionParams,
known_grads={self._q_valsActA: self._action_grad_shared}),
print("Action grads: ", self._action_mean_grads[0])
## When passing in gradients it needs to be a proper list of gradient expressions
self._action_mean_grads = list(self._action_mean_grads[0])
self._actionGRADUpdates = lasagne.updates.adam(
self._action_mean_grads,
self._actionParams,
self._learning_rate,
beta1=0.9,
beta2=0.9,
epsilon=self._rms_epsilon)
self._actGradGivens = {
self._model.getStateSymbolicVariable(): self._model.getStates(),
}
self._actor_regularization = (
self._regularization_weight *
lasagne.regularization.regularize_network_params(
self._model.getActorNetwork(), lasagne.regularization.l2))
### update Actor wrt to Q function
"""
inputs_1_ = {
self._model.getStateSymbolicVariable(): self._model.getStates(),
self._q_valsActA: self._model.getActions()
}
q = self._model.getCriticNetwork()(self._model.getStateSymbolicVariable(), self._q_valsActA)
self._q_valsA_ = lasagne.layers.get_output(self._model.getCriticNetwork(), inputs_1_)
# self._q_valsA_ = lasagne.layers.get_output(self._model.getCriticNetwork(), self._q_valsActA)
self._q_val2 = theano.function([self._model.getStateSymbolicVariable()], self._q_valsA_)
self._actionUpdates = lasagne.updates.adam(-T.mean(self._q_valsA_), self._actionParams,
self._learning_rate, beta1=0.9, beta2=0.9, epsilon=self._rms_epsilon)
"""
## Compute on-policy policy gradient
self._prob = likelihood(self._model.getActionSymbolicVariable(),
self._q_valsActA, self._q_valsActASTD,
self._action_length)
### How should this work if the target network is very odd, as in not a slightly outdated copy.
self._prob_target = likelihood(self._model.getActionSymbolicVariable(),
self._q_valsActTarget_State,
self._q_valsActTargetSTD,
self._action_length)
## This does the sum already
self._r = (self._prob / self._prob_target)
self._actLoss_ = theano.tensor.elemwise.Elemwise(theano.scalar.mul)(
(self._r), self._Advantage)
ppo_epsilon = self.getSettings()['kl_divergence_threshold']
self._actLoss_2 = theano.tensor.elemwise.Elemwise(theano.scalar.mul)(
(theano.tensor.clip(self._r, 1.0 - ppo_epsilon,
1 + ppo_epsilon), self._Advantage))
self._actLoss_ = theano.tensor.minimum((self._actLoss_),
(self._actLoss_2))
self._actLoss = (
(T.mean(self._actLoss_))) + -self._actor_regularization
self._policy_grad = T.grad(-1.0 * self._actLoss, self._actionParams)
self._policy_grad = lasagne.updates.total_norm_constraint(
self._policy_grad, 5)
if (self.getSettings()['optimizer'] == 'rmsprop'):
self._actionUpdates = lasagne.updates.rmsprop(
self._policy_grad, self._actionParams, self._learning_rate,
self._rho, self._rms_epsilon)
elif (self.getSettings()['optimizer'] == 'momentum'):
self._actionUpdates = lasagne.updates.momentum(self._policy_grad,
self._actionParams,
self._learning_rate,
momentum=self._rho)
elif (self.getSettings()['optimizer'] == 'adam'):
self._actionUpdates = lasagne.updates.adam(self._policy_grad,
self._actionParams,
self._learning_rate,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
self._qprop_loss = self._actLoss + T.mean(
(self._QProp_N * self._q_func))
self._policy_grad_loss = self._actLoss
# if ('train_extra_value_function' in self.getSettings() and (self.getSettings()['train_extra_value_function'] == True)):
self._valsA = lasagne.layers.get_output(
self._model._value_function,
self._model.getStateSymbolicVariable(),
deterministic=True)
self._valsA_drop = lasagne.layers.get_output(
self._model._value_function,
self._model.getStateSymbolicVariable(),
deterministic=False)
self._valsNextState = lasagne.layers.get_output(
self._model._value_function,
self._model.getResultStateSymbolicVariable(),
deterministic=True)
self._valsTargetNextState = lasagne.layers.get_output(
self._modelTarget._value_function,
self._model.getResultStateSymbolicVariable(),
deterministic=True)
self._valsTarget = lasagne.layers.get_output(
self._modelTarget._value_function,
self._model.getStateSymbolicVariable(),
deterministic=True)
self._v_target = self._model.getRewardSymbolicVariable() + (
self._discount_factor * self._valsTargetNextState)
self._v_diff = self._v_target - self._valsA
loss_v = T.pow(self._v_diff, 2)
self._v_loss = T.mean(loss_v)
self._params_value = lasagne.layers.helper.get_all_params(
self._model._value_function)
self._givens_value = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getResultStateSymbolicVariable():
self._model.getResultStates(),
self._model.getRewardSymbolicVariable():
self._model.getRewards(),
}
self._value_regularization = (
self._critic_regularization_weight *
lasagne.regularization.regularize_network_params(
self._model._value_function, lasagne.regularization.l2))
self._value_grad = T.grad(self._v_loss + self._value_regularization,
self._params_value)
print("Optimizing Value Function with ",
self.getSettings()['optimizer'], " method")
self._updates_value = lasagne.updates.adam(self._value_grad,
self._params_value,
self._critic_learning_rate,
beta1=0.9,
beta2=0.9,
epsilon=self._rms_epsilon)
self._actGivens_PPO = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
# self._model.getResultStateSymbolicVariable(): self._model.getResultStates(),
# self._model.getRewardSymbolicVariable(): self._model.getRewards(),
self._model.getActionSymbolicVariable():
self._model.getActions(),
# self._NotFallen: self._NotFallen_shared,
self._Advantage:
self._advantage_shared,
# self._KL_Weight: self._kl_weight_shared
}
QProp.compile(self)
def compile(self):
## For training the Q function
self._train = theano.function([], [self._loss, self._q_func],
updates=self._updates_,
givens=self._givens_)
## For training the on-policy value function.
self._train_value = theano.function([], [self._v_loss, self._valsA],
updates=self._updates_value,
givens=self._givens_value)
## Train the policy by backpropgating action gradients
self._trainActionGRAD = theano.function(
[], [],
updates=self._actionGRADUpdates,
givens=self._actGradGivens)
## Train the Policy directly using q function.
# self._trainAction = theano.function([self._model.getStateSymbolicVariable()], [self._q_valsA_], updates=self._actionUpdates)
## Train using PPO like method
self._trainActor = theano.function([
self._model.getStateSymbolicVariable(),
self._model.getActionSymbolicVariable(), self._Advantage
], [self._actLoss],
updates=self._actionUpdates)
## Get the
# self._q_val2 = theano.function([self._model.getStateSymbolicVariable(),
# self._model.getActionSymbolicVariable()], self._q_valsA_)
self._q_val = theano.function(
[],
self._q_func,
givens={
self._model.getStateSymbolicVariable():
self._model.getStates(),
self._model.getActionSymbolicVariable():
self._model.getActions()
})
self._q_val_Target = theano.function(
[ #self._model.getStateSymbolicVariable(),
#self._model.getActionSymbolicVariable()
],
self._q_valsB_,
givens={
self._modelTarget.getStateSymbolicVariable():
self._model.getResultStates(),
self._modelTarget.getActionSymbolicVariable():
self._model.getActions()
})
self._q_action = theano.function(
[],
self._q_valsActA,
givens={
self._model.getStateSymbolicVariable():
self._model.getStates()
})
self._action_Target = theano.function(
[],
self._q_valsActTarget,
givens={
self._model.getResultStateSymbolicVariable():
self._model.getResultStates()
})
self._bellman_error2 = theano.function(inputs=[],
outputs=self._diff,
allow_input_downcast=True,
givens=self._givens_)
self._get_action_grad = theano.function(
[],
outputs=lasagne.updates.get_or_compute_grads(
T.mean(self._q_func),
[self._model._actionInputVar] + self._params),
allow_input_downcast=True,
givens=self._givens_grad)
self._givens_QProp = {
self._model.getStateSymbolicVariable(): self._model.getStates(),
self._model.getActionSymbolicVariable(): self._model.getActions(),
self._Advantage: self._advantage_shared,
self._QProp_N: self._QProp_N_shared
}
self._get_Qprop_action_grad = theano.function(
[],
outputs=lasagne.updates.get_or_compute_grads(
T.mean(self._qprop_loss),
[self._model._actionInputVar] + self._params),
allow_input_downcast=True,
givens=self._givens_QProp)
# self._get_Qprop_action_grad = theano.function([], outputs=lasagne.updates.get_or_compute_grads(self._policy_grad_loss, [self._model._actionInputVar] + self._params), allow_input_downcast=True, givens=self._actGivens_PPO)
self._vals_extra = theano.function(
[],
outputs=self._valsA,
allow_input_downcast=True,
givens={
self._model.getStateSymbolicVariable():
self._model.getStates()
})
if ('train_extra_value_function' in self.getSettings()
and (self.getSettings()['train_extra_value_function'])):
self._givens_grad = {
self._model.getStateSymbolicVariable():
self._model.getStates(),
}
self._get_state_grad = theano.function(
[],
outputs=lasagne.updates.get_or_compute_grads(
T.mean(self._valsA),
[self._model._stateInputVar] + self._params_value),
allow_input_downcast=True,
givens=self._givens_grad)
else:
self._get_state_grad = theano.function(
[],
outputs=lasagne.updates.get_or_compute_grads(
T.mean(self._q_func),
[self._model._stateInputVar] + self._params),
allow_input_downcast=True,
givens=self._givens_grad)
def getGrads(self, states, actions=None, alreadyNormed=False):
"""
The states should be normalized
"""
# self.setData(states, actions, rewards, result_states)
if (alreadyNormed == False):
states = norm_state(states, self._state_bounds)
states = np.array(states, dtype=theano.config.floatX)
self._model.setStates(states)
if (actions is None):
actions = self.predict_batch(states)
self._model.setActions(actions)
return self._get_state_grad()
def getActionGrads(self, states, actions=None, alreadyNormed=False):
"""
The states should be normalized
"""
# self.setData(states, actions, rewards, result_states)
if (alreadyNormed == False):
states = norm_state(states, self._state_bounds)
states = np.array(states, dtype=theano.config.floatX)
self._model.setStates(states)
if (actions is None):
actions = self.predict_batch(states)
self._model.setActions(actions)
if (self._use_basic_polcy_grad):
return self._get_action_grad()
else:
return self._get_Qprop_action_grad()
def updateTargetModel(self):
if (self.getSettings()["print_levels"][self.getSettings(
)["print_level"]] >= self.getSettings()["print_levels"]['train']):
print("Updating target Model")
"""
Target model updates
"""
# return
## I guess it is okay to lerp the entire network even though we only really want to
## lerp the value function part of the networks, the target policy is not used for anythings
all_paramsA = lasagne.layers.helper.get_all_param_values(
self._model.getCriticNetwork())
all_paramsB = lasagne.layers.helper.get_all_param_values(
self._modelTarget.getCriticNetwork())
all_paramsActA = lasagne.layers.helper.get_all_param_values(
self._model.getActorNetwork())
all_paramsActB = lasagne.layers.helper.get_all_param_values(
self._modelTarget.getActorNetwork())
lerp_weight = self.getSettings()['target_net_interp_weight']
all_params = []
for paramsA, paramsB in zip(all_paramsA, all_paramsB):
params = (lerp_weight * paramsA) + ((1.0 - lerp_weight) * paramsB)
all_params.append(params)
all_paramsAct = []
for paramsA, paramsB in zip(all_paramsActA, all_paramsActB):
params = (lerp_weight * paramsA) + ((1.0 - lerp_weight) * paramsB)
all_paramsAct.append(params)
lasagne.layers.helper.set_all_param_values(
self._modelTarget.getCriticNetwork(), all_params)
lasagne.layers.helper.set_all_param_values(
self._modelTarget.getActorNetwork(), all_paramsAct)
def updateTargetModelValue(self):
if (self.getSettings()["print_levels"][self.getSettings(
)["print_level"]] >= self.getSettings()["print_levels"]['train']):
print("Updating MBAE target Model")
"""
Target model updates
"""
all_paramsA = lasagne.layers.helper.get_all_param_values(
self._model._value_function)
lasagne.layers.helper.set_all_param_values(
self._modelTarget._value_function, all_paramsA)
all_paramsActA = lasagne.layers.helper.get_all_param_values(
self._model.getActorNetwork())
lasagne.layers.helper.set_all_param_values(
self._modelTarget2.getActorNetwork(), all_paramsActA)
def updateTargetModel(self):
if (self.getSettings()["print_levels"][self.getSettings(
)["print_level"]] >= self.getSettings()["print_levels"]['train']):
print("Updating target Model")
"""
Target model updates
"""
# return
## I guess it is okay to lerp the entire network even though we only really want to
## lerp the value function part of the networks, the target policy is not used for anythings
all_paramsA = self._model.getCriticNetwork().get_weights()
all_paramsB = self._modelTarget.getCriticNetwork().get_weights()
if ('target_net_interp_weight' in self.getSettings()):
lerp_weight = self.getSettings()['target_net_interp_weight']
else:
lerp_weight = 0.001
# vals = lasagne.layers.helper.get_all_param_values(self._l_outActA)
all_params = []
for paramsA, paramsB in zip(all_paramsA, all_paramsB):
# print ("paramsA: " + str(paramsA))
# print ("paramsB: " + str(paramsB))
params = (lerp_weight * paramsA) + ((1.0 - lerp_weight) * paramsB)
all_params.append(params)
self._modelTarget.getCriticNetwork().set_weights(all_params)
all_paramsA_Act = self._model.getActorNetwork().get_weights()
all_paramsB_Act = self._modelTarget.getActorNetwork().get_weights()
all_params = []
for paramsA, paramsB in zip(all_paramsA_Act, all_paramsB_Act):
# print ("paramsA: " + str(paramsA))
# print ("paramsB: " + str(paramsB))
params = (lerp_weight * paramsA) + ((1.0 - lerp_weight) * paramsB)
all_params.append(params)
self._modelTarget.getActorNetwork().set_weights(all_params)
def getNetworkParameters(self):
params = []
params.append(
lasagne.layers.helper.get_all_param_values(
self._model.getCriticNetwork()))
params.append(
lasagne.layers.helper.get_all_param_values(
self._model.getActorNetwork()))
params.append(
lasagne.layers.helper.get_all_param_values(
self._modelTarget.getCriticNetwork()))
params.append(
lasagne.layers.helper.get_all_param_values(
self._modelTarget.getActorNetwork()))
params.append(
lasagne.layers.helper.get_all_param_values(
self._model._value_function))
params.append(
lasagne.layers.helper.get_all_param_values(
self._modelTarget._value_function))
params.append(
lasagne.layers.helper.get_all_param_values(
self._modelTarget2.getActorNetwork()))
return params
def setNetworkParameters(self, params):
lasagne.layers.helper.set_all_param_values(
self._model.getCriticNetwork(), params[0])
lasagne.layers.helper.set_all_param_values(
self._model.getActorNetwork(), params[1])
lasagne.layers.helper.set_all_param_values(
self._modelTarget.getCriticNetwork(), params[2])
lasagne.layers.helper.set_all_param_values(
self._modelTarget.getActorNetwork(), params[3])
lasagne.layers.helper.set_all_param_values(self._model._value_function,
params[4])
lasagne.layers.helper.set_all_param_values(
self._modelTarget._value_function, params[5])
lasagne.layers.helper.set_all_param_values(
self._modelTarget2.getActorNetwork(), params[6])
def setData(self, states, actions, rewards, result_states, fallen):
self._model.setStates(states)
self._model.setResultStates(result_states)
self._model.setActions(actions)
self._model.setRewards(rewards)
self._modelTarget.setStates(states)
self._modelTarget.setResultStates(result_states)
self._modelTarget.setActions(actions)
self._modelTarget.setRewards(rewards)
self._fallen_shared.set_value(fallen)
def trainOnPolicyCritic(self, states, actions, rewards, result_states,
falls):
"""
Train an on policy value function for the policy
The incoming data should not be normalized.
"""
self.setData(states, actions, rewards, result_states, falls)
loss_v, _ = self._train_value()
print("MBAE Value function loss: ", loss_v)
def trainCritic(self, states, actions, rewards, result_states, falls):
self.setData(states, actions, rewards, result_states, falls)
self._updates += 1
## Compute actions for TargetNet
target_actions = self._action_Target()
self.setData(states, target_actions, rewards, result_states, falls)
## Get next q value
q_vals_b = self._q_val_Target()
## Compute target values
target_tmp_ = rewards + ((self._discount_factor * q_vals_b))
self.setData(states, actions, rewards, result_states, falls)
self._tmp_target_shared.set_value(target_tmp_)
loss, _ = self._train()
# self.updateTargetModel()
return loss
def trainActor(self,
states,
actions,
rewards,
result_states,
falls,
advantage,
exp_actions,
forwardDynamicsModel=None):
if ((self._updates % self._weight_update_steps) == 0):
self.updateTargetModelValue()
self._updates += 1
self.setData(states, actions, rewards, result_states, falls)
if (self.getSettings()["print_levels"][self.getSettings(
)["print_level"]] >= self.getSettings()["print_levels"]['debug']):
print(
"values: ",
np.mean(self._q_val() *
(1.0 / (1.0 - self.getSettings()['discount_factor']))),
" std: ",
np.std(self._q_val() *
(1.0 / (1.0 - self.getSettings()['discount_factor']))))
print("Rewards: ", np.mean(rewards), " std: ", np.std(rewards),
" shape: ",
np.array(rewards).shape)
advantage = advantage * (1.0 - self._discount_factor)
### Get Q value of sampled actions
sampled_q = self._q_val()
### Get Q value of policy
policy_mean = self.predict_batch(states)
self.setData(states, policy_mean, rewards, result_states, falls)
true_q = self._q_val()
### From Q-prop paper, compute adaptive control variate.
cov = advantage * true_q
# var = true_q * true_q
# n = cov / var
### practical implementation n = 1 when cov > 0, otherwise 0
n = (np.sign(cov) + 1.0) / 2.0
n = np.zeros_like(n)
advantage = (advantage - (n * (sampled_q - true_q)))
std = np.std(advantage)
mean = np.mean(advantage)
if ('advantage_scaling' in self.getSettings()
and (self.getSettings()['advantage_scaling'] != False)):
std = std / self.getSettings()['advantage_scaling']
mean = 0.0
advantage = (advantage - mean) / std
if (self._use_basic_polcy_grad):
loss = 0
action_grads = self.getActionGrads(states,
policy_mean,
alreadyNormed=True)[0]
### Get Advantage Action Gradients
action_diff = (actions - policy_mean)
# print ("advantage ", advantage)
if (self.getSettings()["print_levels"][self.getSettings(
)["print_level"]] >= self.getSettings()["print_levels"]['train']):
print("Q-prop n mean: ", np.mean(n), " std: ", np.std(n))
print("Advantage mean: ", np.mean(advantage), " std: ",
np.std(advantage))
print("sampled_q mean: ", np.mean(sampled_q), " std: ",
np.std(sampled_q))
print("true_q mean: ", np.mean(true_q), " std: ",
np.std(true_q))
print("Policy mean: ", np.mean(self._q_action(), axis=0))
# print ("Mean learned advantage: ", np.mean(sampled_q - true_q))
# print ("Mean advantage: " , np.mean(advantage))
action_gra = action_diff * (advantage)
# action_grads = action_gra + ( n * action_grads )
else:
loss = 0
self._advantage_shared.set_value(advantage)
self._QProp_N_shared.set_value(n)
action_grads = self.getActionGrads(states,
policy_mean,
alreadyNormed=True)[0]
action_grads = action_grads
if (self.getSettings()["print_levels"][self.getSettings(
)["print_level"]] >= self.getSettings()["print_levels"]['train']):
print("Q-prop n mean: ", np.mean(n), " std: ", np.std(n))
print("Advantage mean: ", np.mean(advantage), " std: ",
np.std(advantage))
print("sampled_q mean: ", np.mean(sampled_q), " std: ",
np.std(sampled_q))
print("true_q mean: ", np.mean(true_q), " std: ",
np.std(true_q))
print("Policy mean: ", np.mean(self._q_action(), axis=0))
"""
From DEEP REINFORCEMENT LEARNING IN PARAMETERIZED ACTION SPACE
Hausknecht, Matthew and Stone, Peter
actions.shape == action_grads.shape
"""
use_parameter_grad_inversion = False
if (use_parameter_grad_inversion):
for i in range(action_grads.shape[0]):
for j in range(action_grads.shape[1]):
if (action_grads[i, j] > 0):
inversion = (1.0 - actions[i, j]) / 2.0
else:
inversion = (actions[i, j] - (-1.0)) / 2.0
action_grads[i, j] = action_grads[i, j] * inversion
if (self.getSettings()["print_levels"][self.getSettings(
)["print_level"]] >= self.getSettings()["print_levels"]['debug']):
# print("Actions mean: ", np.mean(actions, axis=0))
print("Policy mean: ", np.mean(self._q_action(), axis=0))
# print("Actions std: ", np.mean(np.sqrt( (np.square(np.abs(actions - np.mean(actions, axis=0))))/1.0), axis=0) )
# print("Actions std: ", np.std((actions - self._q_action()), axis=0) )
# print("Actions std: ", np.std((actions), axis=0) )
# print("Policy std: ", np.mean(self._q_action_std(), axis=0))
# print("Mean Next State Grad grad: ", np.mean(next_state_grads, axis=0), " std ", np.std(next_state_grads, axis=0))
print("Mean action grad: ", np.mean(action_grads, axis=0), " std ",
np.std(action_grads, axis=0))
## Set data for gradient
self._model.setStates(states)
self._modelTarget.setStates(states)
## Why the -1.0??
## Because the SGD method is always performing MINIMIZATION!!
# self._action_grad_shared.set_value(-1.0*action_grads)
# self._trainActionGRAD()
self.setData(states, actions, rewards, result_states, falls)
self._advantage_shared.set_value(advantage)
loss = self._trainActor(states, actions, advantage)
# loss = self._trainActor()
return loss
def train(self, states, actions, rewards, result_states):
loss = self.trainCritic(states, actions, rewards, result_states)
lossActor = self.trainActor(states, actions, rewards, result_states)
return loss
def q_value(self, state):
"""
For returning a vector of q values, state should NOT be normalized
"""
# states = np.zeros((self._batch_size, self._state_length), dtype=theano.config.floatX)
# states[0, ...] = state
state = norm_state(state, self._state_bounds)
state = np.array(state, dtype=theano.config.floatX)
self._model.setStates(state)
self._modelTarget.setStates(state)
action = self._q_action()
self._model.setActions(action)
self._modelTarget.setActions(action)
if ('train_extra_value_function' in self.getSettings() and
(self.getSettings()['train_extra_value_function'] == True)):
q_vals = self._vals_extra()
else:
q_vals = self._q_val()
if (('disable_parameter_scaling' in self._settings)
and (self._settings['disable_parameter_scaling'])):
return scale_reward(q_vals, self.getRewardBounds()) * (
1.0 / (1.0 - self.getSettings()['discount_factor']))
# return (self._q_val())[0]
else:
return scale_reward(q_vals, self.getRewardBounds()) * (
1.0 / (1.0 - self.getSettings()['discount_factor']))
# return self._q_valTarget()[0]
# return self._q_val()[0]
def q_values(self, state):
"""
For returning a vector of q values, state should already be normalized
"""
state = norm_state(state, self._state_bounds)
state = np.array(state, dtype=theano.config.floatX)
self._model.setStates(state)
self._modelTarget.setStates(state)
action = self._q_action()
self._model.setActions(action)
self._modelTarget.setActions(action)
if ('train_extra_value_function' in self.getSettings() and
(self.getSettings()['train_extra_value_function'] == True)):
q_vals = self._vals_extra()
else:
q_vals = self._q_val()
if (('disable_parameter_scaling' in self._settings)
and (self._settings['disable_parameter_scaling'])):
return scale_reward(q_vals, self.getRewardBounds()) * (
1.0 / (1.0 - self.getSettings()['discount_factor']))
# return (self._q_val())[0]
else:
return scale_reward(q_vals, self.getRewardBounds()) * (
1.0 / (1.0 - self.getSettings()['discount_factor']))
# return self._q_valTarget()
# return self._q_val()
def setStateBounds(self, state_bounds):
super(QProp, self).setStateBounds(state_bounds)
self._experience.setStateBounds(copy.deepcopy(self.getStateBounds()))
def setActionBounds(self, action_bounds):
super(QProp, self).setActionBounds(action_bounds)
self._experience.setActionBounds(copy.deepcopy(self.getActionBounds()))
def setRewardBounds(self, reward_bounds):
super(QProp, self).setRewardBounds(reward_bounds)
self._experience.setRewardBounds(copy.deepcopy(self.getRewardBounds()))
def _q_action_std(self):
ones = np.ones((self._model.getStateValues().shape[0],
len(self.getActionBounds()[0])))
return np.array(self.getSettings()["exploration_rate"] * ones)