Ejemplo n.º 1
0
    def __init__(self, game, paramOnly=False, nReps=1, record=False):
        """Initializes task environment
  
    Args:
      game - (string) - dict key of task to be solved (see domain/config.py)
  
    Optional:
      paramOnly - (bool)  - only load parameters instead of launching task?
      nReps     - (nReps) - number of trials to get average fitness
    """
        # Network properties
        self.nInput = game.input_size
        self.nOutput = game.output_size
        self.actRange = game.h_act
        self.absWCap = game.weightCap
        self.layers = game.layers
        self.activations = np.r_[np.full(1, 1), game.i_act, game.o_act]

        # Environment
        self.nReps = nReps
        self.maxEpisodeLength = game.max_episode_length
        self.actSelect = game.actionSelect
        if not paramOnly:
            if record:
                env_to_wrap = make_env(game.env_name)
                self.env = Monitor(env_to_wrap, "trial_recording/", force=True)
            else:
                self.env = make_env(game.env_name)

        # Special needs...
        self.needsClosed = (game.env_name.startswith("CartPoleSwingUp"))
Ejemplo n.º 2
0
    def __init__(self, game, encoder, max_features, paramOnly=False, nReps=1):
        """Initializes task environment
  
    Args:
      game - (string) - dict key of task to be solved (see domain/config.py)
  
    Optional:
      paramOnly - (bool)  - only load parameters instead of launching task?
      nReps     - (nReps) - number of trials to get average fitness
    """
        # Network properties
        if encoder == 'ascii' or encoder == 'count':
            self.nInput = max_features
            i_act = np.full(max_features, 1)
        elif encoder == 'lstm':
            self.nInput = 4096
            i_act = np.full(4096, 1)
        else:
            self.nInput = game.input_size
            i_act = game.i_act

        self.nOutput = game.output_size
        self.actRange = game.h_act
        self.absWCap = game.weightCap
        self.layers = game.layers
        self.activations = np.r_[np.full(1, 1), i_act, game.o_act]

        # Environment
        self.nReps = nReps
        self.maxEpisodeLength = game.max_episode_length
        self.actSelect = game.actionSelect
        if not paramOnly:
            self.env = make_env(game.env_name, encoder, max_features)

        # Special needs...
        self.needsClosed = (game.env_name.startswith("CartPoleSwingUp"))