actor_network.py 文件源码

def create_network(self,state_dim,action_dim,scope):
        with tf.variable_scope(scope,reuse=False) as s:

            state_input = tf.placeholder("float",[None,None,state_dim])

            # creating the recurrent part
            lstm_cell=rnn.BasicLSTMCell(LSTM_HIDDEN_UNIT)
            lstm_output,lstm_state=tf.nn.dynamic_rnn(cell=lstm_cell,inputs=state_input,dtype=tf.float32)
            W3 = tf.Variable(tf.random_uniform([lstm_cell.state_size,action_dim],-3e-3,3e-3))
            b3 = tf.Variable(tf.random_uniform([action_dim],-3e-3,3e-3))

            action_output = tf.tanh(tf.matmul(lstm_state,W3) + b3)

            net = [v for v in tf.trainable_variables() if scope in v.name]

        return state_input,action_output,net