yapf

class TicTacToePolicy(dc.rl.Policy):

  def create_layers(self, state, **kwargs):

    d1 = Flatten(in_layers=state)

    d2 = Dense(

    return {'action_prob': probs, 'value': value}

def eval_tic_tac_toe(value_weight, num_epoch_rounds=1, games=10 ** 4, rollouts=10 ** 5):

def eval_tic_tac_toe(value_weight,

                     num_epoch_rounds=1,

                     games=10**4,

                     rollouts=10**5):

  """

    Returns the average reward over 1k games after 10k rollouts

    :param value_weight:

  avg_rewards = []

  for j in range(num_epoch_rounds):

        a3c = dc.rl.A3C(env, policy, entropy_weight=0.01, value_weight=value_weight, model_dir=model_dir)

        a3c.optimizer = dc.models.tensorgraph.TFWrapper(tf.train.AdamOptimizer, learning_rate=0.01)

    a3c = dc.rl.A3C(

        env,

        policy,

        entropy_weight=0.01,

        value_weight=value_weight,

        model_dir=model_dir)

    a3c.optimizer = dc.models.tensorgraph.TFWrapper(

        tf.train.AdamOptimizer, learning_rate=0.01)

    try:

      a3c.restore()

    except:

def main():

  value_weight = 6.0

    score = eval_tic_tac_toe(value_weight, num_epoch_rounds=10)

  score = eval_tic_tac_toe(value_weight, num_epoch_rounds=1)

  print(score)

  def create_tensor(self, **kwargs):

    reward, action, prob, value = [layer.out_tensor for layer in self.in_layers]

    log_prob = tf.log(prob + 0.0001)

    prob = prob + np.finfo(np.float32).eps

    log_prob = tf.log(prob)

    policy_loss = -tf.reduce_sum(

        (reward - value) * tf.reduce_sum(action * log_prob))

    value_loss = tf.reduce_sum(tf.square(reward - value))