def setup_models(self):
dqn = self.setup_dqn(trainable=False)
teammates_model = TeamModel(num_teammates=self.num_teammates,
trainable=False)
dqn.load(self.directory, load_memory=False)
teammates_model.load(self.directory, load_memory=False)
return dqn, teammates_model
class LearntPLASTICModel(PLASTICPrior):
def __init__(self, directory, teammates, num_teammates):
self.teammates_model = TeamModel(num_teammates=num_teammates, trainable=False)
self.teammates_model.load(directory, load_memory=False)
super().__init__(teammates, num_teammates)
#################
# PLASTIC Prior #
#################
def policies(self, state):
return self.teammates_model.policies(state)
class LearningPLASTICModel(PLASTICPrior):
def __init__(self, num_teammates):
self.teammates_model = TeamModel(num_teammates, trainable=True)
super().__init__("new", num_teammates)
############################
# Learning Prior Interface #
############################
def reinforce(self, state, joint_actions, reward, next_state, terminal):
info = {}
datapoint = state, joint_actions, reward, next_state, terminal
info["team model"] = self.teammates_model.replay_fit(datapoint)
return info
def save(self, directory):
prior_dir = f"{directory}/{self.name}"
mkdir(prior_dir)
self.teammates_model.save(prior_dir)
def load(self, directory):
prior_dir = f"{directory}/{self.name}"
self.teammates_model.load(prior_dir)
#################
# PLASTIC Prior #
#################
def policies(self, state):
return self.teammates_model.policies(state)
class PlasticPolicy:
def __init__(self, num_teammates):
self._num_teammates = num_teammates
self.total_timesteps = 0
self.dqn, self.teammates_model = self.setup_models()
@staticmethod
def setup_dqn(trainable):
dqn = DQN(trainable, config.LEARNING_RATE, config.DQN_DISCOUNT_FACTOR,
ReplayMemoryModel.parse_layer_blueprints(cfg["layers"]),
config.REPLAY_MIN_BATCH, config.REPLAY_MEMORY_SIZE)
return dqn
def reinforce(self, state, joint_actions, reward, next_state, terminal):
info = {}
datapoint = state, joint_actions, reward, next_state, terminal
self.total_timesteps += 1
info["dqn"] = self.dqn.replay_fit(datapoint)
info["team model"] = self.team_model.replay_fit(datapoint)
return info
def save(self, directory):
prior_dir = f"{directory}/{self.name}"
mkdir(prior_dir)
self.dqn.save(prior_dir)
self.team_model.save(prior_dir)
def load(self, directory):
prior_dir = f"{directory}/{self.name}"
self.dqn.load(prior_dir)
self.team_model.load(prior_dir)
##################
# PLASTIC Policy #
##################
def setup_models(self):
self.dqn = self.setup_dqn(trainable=True)
self.team_model = TeamModel(self._num_teammates, trainable=True)
return self.dqn, self.team_model
#################
# PLASTIC Prior #
#################
def policies(self, state):
return self.team_model.policies(state)
class LearningPLASTICPolicy(PLASTICPolicy):
def __init__(self, num_teammates):
self._num_teammates = num_teammates
self.total_timesteps = 0
super().__init__("new", num_teammates)
############################
# Learning Prior Interface #
############################
def reinforce(self, state, joint_actions, reward, next_state, terminal):
info = {}
datapoint = state, joint_actions, reward, next_state, terminal
self.total_timesteps += 1
info["dqn"] = self.dqn.replay_fit(datapoint)
info["team model"] = self.team_model.replay_fit(datapoint)
return info
def save(self, directory):
prior_dir = f"{directory}/{self.name}"
mkdir(prior_dir)
self.dqn.save(prior_dir)
self.team_model.save(prior_dir)
def load(self, directory):
prior_dir = f"{directory}/{self.name}"
self.dqn.load(prior_dir)
self.team_model.load(prior_dir)
##################
# PLASTIC Policy #
##################
def setup_models(self):
self.dqn = self.setup_dqn(trainable=True)
self.team_model = TeamModel(self._num_teammates, trainable=True)
return self.dqn, self.team_model
#################
# PLASTIC Prior #
#################
def policies(self, state):
return self.team_model.policies(state)
def setup_models(self):
self.dqn = self.setup_dqn(trainable=True)
self.team_model = TeamModel(self._num_teammates, trainable=True)
return self.dqn, self.team_model
def __init__(self, num_teammates):
self.teammates_model = TeamModel(num_teammates, trainable=True)
super().__init__("new", num_teammates)
def __init__(self, directory, teammates, num_teammates):
self.teammates_model = TeamModel(num_teammates=num_teammates, trainable=False)
self.teammates_model.load(directory, load_memory=False)
super().__init__(teammates, num_teammates)