def __init__(self):
AgentBase.__init__(self)
self.ClassAct = ShareBiConv
self.ClassCri = self.ClassAct
self.if_use_cri_target = True
self.if_use_act_target = True
self.obj_critic = (-np.log(0.5))**0.5 # for reliable_lambda
def __init__(self):
AgentBase.__init__(self)
self.ClassAct = ActorBiConv
self.ClassCri = CriticBiConv
self.if_use_cri_target = False
self.if_use_act_target = False
self.explore_noise = 2**-8
self.obj_critic = (-np.log(0.5))**0.5 # for reliable_lambda
def __init__(self):
AgentBase.__init__(self)
self.ClassAct = Actor
self.ClassCri = Critic
self.if_use_cri_target = True
self.if_use_act_target = True
self.explore_noise = 0.3 # explore noise of action (OrnsteinUhlenbeckNoise)
self.ou_noise = None
def __init__(self):
AgentBase.__init__(self)
self.ClassAct = Actor
self.ClassCri = CriticTwin
self.if_use_cri_target = True
self.if_use_act_target = True
self.explore_noise = 0.1 # standard deviation of exploration noise
self.policy_noise = 0.2 # standard deviation of policy noise
self.update_freq = 2 # delay update frequency
def __init__(self):
AgentBase.__init__(self)
self.ClassAct = ActorPPO
self.ClassCri = CriticPPO
self.if_off_policy = False
self.ratio_clip = 0.2 # could be 0.00 ~ 0.50 ratio.clamp(1 - clip, 1 + clip)
self.lambda_entropy = 0.02 # could be 0.00~0.10
self.lambda_a_value = 1.00 # could be 0.25~8.00, the lambda of advantage value
self.lambda_gae_adv = 0.98 # could be 0.95~0.99, GAE (Generalized Advantage Estimation. ICLR.2016.)
self.get_reward_sum = None # self.get_reward_sum_gae if if_use_gae else self.get_reward_sum_raw
def __init__(self):
AgentBase.__init__(self)
self.ClassCri = CriticTwin
self.ClassAct = ActorSAC
self.if_use_cri_target = True
self.if_use_act_target = False
self.alpha_log = None
self.alpha_optim = None
self.target_entropy = None
self.obj_critic = (-np.log(0.5))**0.5 # for reliable_lambda
def __init__(self):
AgentBase.__init__(self)
self.ClassCri = DiscreteCriSAC
self.ClassAct = DiscreteActSAC
self.train_reward = []
self.if_use_cri_target = True
self.if_use_act_target = False
self.trajectory_list = []
self.alpha_log = None
self.alpha_optim = None
self.target_entropy = None
self.obj_critic = (-np.log(0.5))**0.5 # for reliable_lambda
self.train_iteraion = 0
def __init__(self):
AgentBase.__init__(self)
self.ClassCri = None # self.ClassCri = QNetDuel if self.if_use_dueling else QNet
self.if_use_dueling = True # self.ClassCri = QNetDuel if self.if_use_dueling else QNet
self.explore_rate = 0.25 # the probability of choosing action randomly in epsilon-greedy