capacities.py 文件源码

def tabular_learning_with_lr(init_lr, decay_steps, Qs_t, states_t, actions_t, targets):
    reusing_scope = tf.get_variable_scope().reuse

    state_action_pairs = tf.stack([states_t, actions_t], 1)
    estimates = tf.gather_nd(Qs_t, state_action_pairs)
    err_estimates = targets - estimates
    loss = tf.reduce_mean(err_estimates)

    global_step = tf.Variable(0, trainable=False, name="global_step", collections=[tf.GraphKeys.GLOBAL_STEP, tf.GraphKeys.GLOBAL_VARIABLES])
    lr = tf.train.exponential_decay(tf.constant(init_lr, dtype=tf.float32), global_step, decay_steps, 0.5, staircase=True)
    if reusing_scope is False:
        tf.summary.scalar('lr', lr)
    inc_global_step = global_step.assign_add(1)
    with tf.control_dependencies([inc_global_step]):
        updates = lr * err_estimates
        train_op = tf.scatter_nd_add(Qs_t, state_action_pairs, updates)

    return loss, train_op