model.setRewardBounds(reward_bounds)
model.setActionBounds(action_bounds)
arr = list(range(len(states_)))
random.shuffle(arr)
given_actions = []
given_states = []
for i in range(len(states_)):
a = actions[arr[i]]
action_ = np.array([a])
given_actions.append(action_)
state_ = np.array([states_[arr[i]]])
next_state_ = np.array([next_states_[arr[i]]])
given_states.append(state_)
# print "Action: " + str([actions[i]])
experience.insert(state_, action_, next_state_, np.array([1]))
# print ("Added tuple: ", i)
errors = []
for i in range(settings['rounds']):
# print ("Actions: ", _actions)
# print ("States: ", _states)
# (error, lossActor) = model.train(_states, _actions, _result_states, _rewards)
for j in range(1):
_states, _actions, _result_states, _rewards, falls_, advantage, exp_actions__ = experience.get_batch(
batch_size)
error = model.trainCritic(_states, _actions, _result_states,
_rewards)
for j in range(5):
_states, _actions, _result_states, _rewards, falls_, advantage, exp_actions__ = experience.get_batch(
batch_size)
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]])
experience.insert(state_, state_, action_, np.array([0]))
errors = []
for i in range(1000):
_states, _actions, _result_states, _rewards, fals_, _G_ts, advantage = experience.get_batch(
batch_size)
# print ("Actions: ", _actions)
# print ("States: ", _states)
error = model.train(_states, _states, _result_states, _actions)
errors.append(error)
# print "Error: " + str(error)
states = np.linspace(-5.0, 5.0, experience_length)
actionsNoNoise = np.array(map(f, states))
# print ("Eval States: ", np.transpose(np.array([states])))
# states2 = np.transpose(np.repeat([states], 2, axis=0))
# print states2
model = NeuralNetwork(len(state_bounds[0]), len(action_bounds[0]),
state_bounds, action_bounds, settings)
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))
errors = []
for i in range(5000):
_states, _actions, _result_states, _rewards = experience.get_batch(
batch_size)
# print _actions
error = model.train(_states, _actions)
errors.append(error)
# print "Error: " + str(error)
states = np.linspace(-1.0, 6.0, experience_length)
actionsNoNoise = np.array(map(f, states))
predicted_actions = np.array(list(map(model.predict, states)))