def __init__(self, env_space, cmdl):
self.name = "Evaluation"
self.actions = env_space[0]
self.action_no = action_no = self.actions.n
self.cmdl = cmdl
self.epsilon = 0.05
if cmdl.agent_type == "dqn":
self.policy = policy = get_model(cmdl.estimator, 1, cmdl.hist_len,
self.action_no, cmdl.hidden_size)
if self.cmdl.cuda:
self.policy.cuda()
self.policy_evaluation = DeterministicPolicy(policy)
elif cmdl.agent_type == "categorical":
self.policy = policy = get_model(cmdl.estimator,
1,
cmdl.hist_len,
(action_no, cmdl.atoms_no),
hidden_size=cmdl.hidden_size)
if self.cmdl.cuda:
self.policy.cuda()
self.policy_evaluation = CategoricalPolicyEvaluation(policy, cmdl)
print("[%s] Evaluating %s agent." % (self.name, cmdl.agent_type))
self.max_q = -1000
def __init__(self, action_space, cmdl, is_training=True):
DQNAgent.__init__(self, action_space, cmdl, is_training)
self.name = "Categorical_agent"
self.cmdl = cmdl
hist_len, action_no = cmdl.hist_len, self.action_no
self.policy = policy = get_model(cmdl.estimator, 1, hist_len,
(action_no, cmdl.atoms_no),
hidden_size=cmdl.hidden_size)
self.target = target = get_model(cmdl.estimator, 1, hist_len,
(action_no, cmdl.atoms_no),
hidden_size=cmdl.hidden_size)
if self.cmdl.cuda:
self.policy.cuda()
self.target.cuda()
self.policy_evaluation = CategoricalPolicyEvaluation(policy, cmdl)
self.policy_improvement = CategoricalPolicyImprovement(
policy, target, cmdl)
def __init__(self, action_space, cmdl, is_training=True):
DQNAgent.__init__(self, action_space, cmdl, is_training)
self.name = "Categorical_agent"
self.cmdl = cmdl
hist_len, action_no = cmdl.hist_len, self.action_no
self.policy = policy = get_model(cmdl.estimator,
1,
hist_len, (action_no, cmdl.atoms_no),
hidden_size=cmdl.hidden_size)
self.target = target = get_model(cmdl.estimator,
1,
hist_len, (action_no, cmdl.atoms_no),
hidden_size=cmdl.hidden_size)
if self.cmdl.cuda:
self.policy.cuda()
self.target.cuda()
self.policy_evaluation = CategoricalPolicyEvaluation(policy, cmdl)
self.policy_improvement = CategoricalPolicyImprovement(
policy, target, cmdl)
def __init__(self, action_space, cmdl, is_training=True):
BaseAgent.__init__(self, action_space, is_training)
self.name = "DQN_agent"
self.cmdl = cmdl
eps = self.cmdl.epsilon
e_steps = self.cmdl.epsilon_steps
self.policy = policy = get_model(cmdl.estimator, 1, cmdl.hist_len,
self.action_no, cmdl.hidden_size)
self.target = target = get_model(cmdl.estimator, 1, cmdl.hist_len,
self.action_no, cmdl.hidden_size)
if self.cmdl.cuda:
self.policy.cuda()
self.target.cuda()
self.policy_evaluation = DQNEvaluation(policy)
self.policy_improvement = DQNImprovement(policy, target, cmdl)
self.exploration = get_epsilon_schedule("linear", eps, 0.05, e_steps)
self.replay_memory = ReplayMemory(capacity=cmdl.experience_replay)
self.dtype = TorchTypes(cmdl.cuda)
self.max_q = -1000
def __init__(self, action_space, cmdl, is_training=True):
BaseAgent.__init__(self, action_space, is_training)
self.name = "DQN_agent"
self.cmdl = cmdl
eps = self.cmdl.epsilon
e_steps = self.cmdl.epsilon_steps
self.policy = policy = get_model(cmdl.estimator, 1, cmdl.hist_len,
self.action_no, cmdl.hidden_size)
self.target = target = get_model(cmdl.estimator, 1, cmdl.hist_len,
self.action_no, cmdl.hidden_size)
if self.cmdl.cuda:
self.policy.cuda()
self.target.cuda()
self.policy_evaluation = DQNEvaluation(policy)
self.policy_improvement = DQNImprovement(policy, target, cmdl)
self.exploration = get_epsilon_schedule("linear", eps, 0.05, e_steps)
self.replay_memory = ReplayMemory(capacity=cmdl.experience_replay)
self.dtype = TorchTypes(cmdl.cuda)
self.max_q = -1000