ddpg.py 文件源码

def create_critic_network(self, state_size, action_dim):
        """create critic network."""
        print ("[MESSAGE] Build critic network.""")
        S = Input(shape=state_size)
        A = Input(shape=(action_dim,))

        # input
        h_0 = Conv2D(32, (3, 3), padding="same",
                     kernel_regularizer=l2(0.0001),
                     activation="relu")(S)
        h_1 = Conv2D(32, (3, 3), padding="same",
                     kernel_regularizer=l2(0.0001),
                     activation="relu")(h_0)
        h_1 = AveragePooling2D(2, 2)(h_1)
        h_1 = Flatten()(h_1)
        h_1 = Dense(600, activation="relu")(h_1)

        # action
        a_1 = Dense(600, activation="linear")(A)
        h_2 = add([h_1, a_1])
        h_3 = Dense(600, activation="relu")(h_2)
        V = Dense(action_dim, activation="softmax")(h_3)

        model = Model(inputs=[S, A], outputs=V)
        model.compile(loss='categorical_crossentropy',
                      optimizer="adam")
        return model, A, S