knightower.py 文件源码

def surrogate_objective(policy_out):
    """
    Create the surrogate objective for policy gradients.

    Returns actions, rewards, objective.
    """
    actions = tf.placeholder(tf.float32, [None, 2])
    rewards = tf.placeholder(tf.float32, [None, 1])
    objective = tf.tensordot(tf.log(policy_out), actions*rewards, axes=2)
    return actions, rewards, objective