# Same for the discount factor.
agent.attach(
bc.DiscountFactorController(
initial_discount_factor=parameters.discount,
discount_factor_growth=parameters.discount_inc,
discount_factor_max=parameters.discount_max,
periodicity=1))
# As for the discount factor and the learning rate, one can update periodically the parameter of the epsilon-greedy
# policy implemented by the agent. This controllers has a bit more capabilities, as it allows one to choose more
# precisely when to update epsilon: after every X action, episode or epoch. This parameter can also be reset every
# episode or epoch (or never, hence the resetEvery='none').
agent.attach(
bc.EpsilonController(initial_e=parameters.epsilon_start,
e_decays=parameters.epsilon_decay,
e_min=parameters.epsilon_min,
evaluate_on='action',
periodicity=1,
reset_every='none'))
# We wish to discover, among all versions of our neural network (i.e., after every training epoch), which one
# seems to generalize the better, thus which one has the highest validation score. However we also want to keep
# track of a "true generalization score", the "test score". Indeed, what if we overfit the validation score ?
# To achieve these goals, one can use the FindBestController along two InterleavedTestEpochControllers, one for
# each mode (validation and test). It is important that the validationID and testID are the same than the id
# argument of the two InterleavedTestEpochControllers (implementing the validation mode and test mode
# respectively). The FindBestController will dump on disk the validation and test scores for each and every
# network, as well as the structure of the neural network having the best validation score. These dumps can then
# used to plot the evolution of the validation and test scores (see below) or simply recover the resulting neural
# network for your application.
agent.attach(
bc.FindBestController(validationID=MG_two_storages_env.VALIDATION_MODE,
random_state=rng)
agent = NeuralAgent(env,
network,
replay_memory_size=min(
int(args.epochs[0] * args.epochs[1] * 1.1),
100000),
batch_size=32,
random_state=rng)
agent.setDiscountFactor(0.95)
agent.attach(bc.FindBestController(validationID=0,
unique_fname=args.fname))
agent.attach(bc.VerboseController())
agent.attach(bc.TrainerController())
agent.attach(
bc.EpsilonController(initial_e=0.8,
e_decays=args.epochs[0] * args.epochs[1],
e_min=0.2))
agent.attach(
bc.LearningRateController(args.learning_rate[0], args.learning_rate[1],
args.learning_rate[2]))
agent.attach(
bc.InterleavedTestEpochController(epoch_length=1000,
controllers_to_disable=[1, 2, 3, 4]))
elif args.network == 'DDPG':
network = MyACNetwork(environment=env,
batch_size=32,
double_Q=True,
freeze_interval=args.epochs[1],
random_state=rng)
agent = NeuralAgent(
env,
# Same for the discount factor.
agent.attach(bc.DiscountFactorController(
initialDiscountFactor=parameters.discount,
discountFactorGrowth=parameters.discount_inc,
discountFactorMax=parameters.discount_max,
periodicity=1))
# As for the discount factor and the learning rate, one can update periodically the parameter of the epsilon-greedy
# policy implemented by the agent. This controllers has a bit more capabilities, as it allows one to choose more
# precisely when to update epsilon: after every X action, episode or epoch. This parameter can also be reset every
# episode or epoch (or never, hence the resetEvery='none').
agent.attach(bc.EpsilonController(
initialE=parameters.epsilon_start,
eDecays=parameters.epsilon_decay,
eMin=parameters.epsilon_min,
evaluateOn='action',
periodicity=1,
resetEvery='none'))
# We wish to discover, among all versions of our neural network (i.e., after every training epoch), which one
# seems to generalize the better, thus which one has the highest validation score. However we also want to keep
# track of a "true generalization score", the "test score". Indeed, what if we overfit the validation score ?
# To achieve these goals, one can use the FindBestController along two InterleavedTestEpochControllers, one for
# each mode (validation and test). It is important that the validationID and testID are the same than the id
# argument of the two InterleavedTestEpochControllers (implementing the validation mode and test mode
# respectively). The FindBestController will dump on disk the validation and test scores for each and every
# network, as well as the structure of the neural network having the best validation score. These dumps can then
# used to plot the evolution of the validation and test scores (see below) or simply recover the resulting neural
# network for your application.
agent.attach(bc.FindBestController(
def __init__(self,
inputDims,
q_network,
actions,
file='',
replay_memory_size=Parameters.REPLAY_MEMORY_SIZE,
replay_start_size=Parameters.BATCH_SIZE,
batch_size=Parameters.BATCH_SIZE,
random_state=np.random.RandomState(),
exp_priority=0,
batch_type='sequential',
train_policy=None,
test_policy=None,
only_full_history=True,
reward_as_input=False):
CompleteLearner.__init__(self, actions, file)
self.polfile = open(self.file + 'policy.txt', "w")
# if replay_start_size == None:
# replay_start_size = max(inputDims[i][0] for i in range(len(inputDims)))
# elif replay_start_size < max(inputDims[i][0] for i in range(len(inputDims))):
# raise AgentError("Replay_start_size should be greater than the biggest history of a state.")
self._controllers = []
# --- Bind controllers to the agent ---
# For comments, please refer to run_toy_env.py
self.attach(bc.VerboseController(evaluate_on='epoch', periodicity=1))
self.attach(
bc.TrainerController(evaluate_on='action',
periodicity=Parameters.UPDATE_FREQUENCY,
show_episode_avg_V_value=True,
show_avg_Bellman_residual=True))
self.attach(
bc.LearningRateController(
initial_learning_rate=Parameters.LEARNING_RATE,
learning_rate_decay=Parameters.LEARNING_RATE_DECAY,
periodicity=10000))
self.attach(
bc.DiscountFactorController(
initial_discount_factor=Parameters.DISCOUNT,
discount_factor_growth=Parameters.DISCOUNT_INC,
discount_factor_max=Parameters.DISCOUNT_MAX,
periodicity=10000))
self.attach(
bc.EpsilonController(initial_e=Parameters.EPSILON_START,
e_decays=Parameters.EPSILON_DECAY,
e_min=Parameters.EPSILON_MIN,
evaluate_on='action',
periodicity=1000,
reset_every='none'))
# self.attach(bc.InterleavedTestEpochController(
# id=0,
# epoch_length=Parameters.STEPS_PER_TEST,
# controllers_to_disable=[0, 1, 2, 3, 4],
# periodicity=2,
# show_score=True,
# summarize_every=-1))
self.obs = []
self.reward_as_input = reward_as_input
self._network = q_network
self._replay_memory_size = replay_memory_size
self._replay_start_size = replay_start_size # make sure you gather this many observations before learning
self._batch_size = batch_size
self._batch_type = batch_type
self._random_state = random_state
self._exp_priority = exp_priority
self._only_full_history = only_full_history
#inputDimensions, n_actions, observation_type = np.float32, random_state = None, max_size = 1000, batch_type = 'random', only_full_history = True
self._dataset = DataSet(inputDimensions=inputDims,
n_actions=len(actions),
max_size=replay_memory_size,
random_state=random_state,
batch_type=self._batch_type,
only_full_history=self._only_full_history)
self._tmp_dataset = None # Will be created by startTesting() when necessary
self._mode = -1
self._mode_epochs_length = 0
self._total_mode_reward = 0
self._training_loss_averages = []
self._Vs_on_last_episode = []
#self._in_episode = False
self._selected_action = -1
self._state = []
for i in range(len(inputDims)):
self._state.append(np.zeros(inputDims[i], dtype=config.floatX))
if (train_policy == None):
self._train_policy = EpsilonGreedyPolicy(q_network, len(actions),
random_state, 0.05)
else:
self._train_policy = train_policy
if (test_policy == None):
self._test_policy = EpsilonGreedyPolicy(q_network, len(actions),
random_state, 0.)
else:
self._test_policy = test_policy
self.initEpisode()