Merge pull request #640 from peastman/scale

    ]

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

    log_prob = tf.log(prob)

    policy_loss = -tf.reduce_sum(advantage * tf.reduce_sum(action * log_prob))

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

    entropy = -tf.reduce_sum(prob * log_prob)

    policy_loss = -tf.reduce_mean(advantage * tf.reduce_sum(action * log_prob))

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

    entropy = -tf.reduce_mean(tf.reduce_sum(prob * log_prob, axis=1))

    self.out_tensor = policy_loss + self.value_weight * value_loss - self.entropy_weight * entropy

    return self.out_tensor

    a3c = dc.rl.A3C(

        env,

        TestPolicy(),

        max_rollout_length=50,

        max_rollout_length=20,

        optimizer=dc.models.tensorgraph.TFWrapper(

            tf.train.AdamOptimizer, learning_rate=0.0001))

            tf.train.AdamOptimizer, learning_rate=0.001))

    a3c.fit(100000)

    # It should have learned that the expected value is very close to zero, and that the best