reinforce.py 文件源码

def build_network(self):
        # Symbolic variables for observation, action, and advantage
        self.states = tf.placeholder(tf.float32, [None, self.env_runner.nO], name="states")  # Observation
        self.a_n = tf.placeholder(tf.float32, name="a_n")  # Discrete action
        self.adv_n = tf.placeholder(tf.float32, name="adv_n")  # Advantage

        L1 = tf.contrib.layers.fully_connected(
            inputs=self.states,
            num_outputs=self.config["n_hidden_units"],
            activation_fn=tf.tanh,
            weights_initializer=tf.random_normal_initializer(),
            biases_initializer=tf.zeros_initializer())

        self.probs = tf.contrib.layers.fully_connected(
            inputs=L1,
            num_outputs=self.env_runner.nA,
            activation_fn=tf.nn.softmax,
            weights_initializer=tf.random_normal_initializer(),
            biases_initializer=tf.zeros_initializer())

        self.action = tf.squeeze(tf.multinomial(tf.log(self.probs), 1), name="action")