multistate_dqn.py 文件源码

def __init__(self, state_shape, action_num, image_num_per_state,
            model,
            gamma=0.99, # discount factor
            replay_batch_size=32,
            replay_memory_size=5*10**4,
            target_model_update_freq=1,
            max_step=50,
            lr=0.00025,
            clipping=False # if True, ignore reward intensity
            ):
        print("initializing DQN...")
        self.action_num = action_num
        self.image_num_per_state = image_num_per_state
        self.gamma = gamma
        self.replay_batch_size = replay_batch_size
        self.replay_memory_size = replay_memory_size
        self.target_model_update_freq = target_model_update_freq
        self.max_step = max_step
        self.clipping = clipping

        print("Initializing Model...")
        self.model = model
        self.model_target = copy.deepcopy(self.model)

        print("Initializing Optimizer")
        self.optimizer = optimizers.RMSpropGraves(lr=lr, alpha=0.95, momentum=0.0, eps=0.01)
        self.optimizer.setup(self.model)
        self.optimizer.add_hook(chainer.optimizer.GradientClipping(20))

        print("Initializing Replay Buffer...")
        self.dataset = dataset.DataSet(
                max_size=replay_memory_size, max_step=max_step, frame_shape=state_shape, frame_dtype=np.uint8)

        self.xp = model.xp
        self.state_shape = state_shape