tensorflow实现强化学习DDPG算法

基于tf2.x实现DDPG

1. 相关依赖库

• tensorflow==2.7.0
• gym==0.24.0
• python==3.8.0
• 测试运行环境——windows/Mac

2. DDPG代码实现

import tensorflow as tf
from tensorflow import keras
import numpy as np


class DDPG(keras.Model):
    def __init__(self, a_dim, s_dim, a_bound, batch_size=32, tau=0.002, gamma=0.95,
                 a_lr=0.0001, c_lr=0.001, memory_capacity=9000):
        super().__init__()
        self.batch_size = batch_size  # 批量数据
        self.tau = tau   # 滑动平均参数
        self.gamma = gamma   # 回报折扣系数
        self.a_lr = a_lr  # actor学习率
        self.c_lr = c_lr  # critic学习率
        self.memory_capacity = memory_capacity  # 记忆库大小
        self.memory = np.zeros((memory_capacity, s_dim * 2 + a_dim + 1), dtype=np.float32)
        self.pointer = 0  # 记忆库初始大小为0
        self.memory_full = False  # 记忆库是否已经满

        self.a_dim, self.s_dim, self.a_bound = a_dim, s_dim, a_bound[1]  # 动作维度， 状态维度， 动作值的上限

        s = keras.Input(shape=(s_dim,))     # current state
        s_ = keras.Input(shape=(s_dim,))    # next state
        self.actor = self._build_actor(trainable=True, name="a/eval")  # 策略网络
        self.actor_ = self._build_actor(trainable=False, name="a/target")  # 目标策略网络
        self.actor_.set_weights(self.actor.get_weights())
        self.critic = self._build_critic(trainable=True, name="d/eval")  # Q网络
        self.critic_ = self._build_critic(trainable=False, name="d/target")  # 目标Q网络
        self.critic_.set_weights(self.critic.get_weights())
        self.a_opt = keras.optimizers.Adam(self.a_lr)  # 优化器
        self.c_opt = keras.optimizers.Adam(self.c_lr)  # 优化器
        self.mse = keras.losses.MeanSquaredError()  # 均方差损失函数

    def _build_actor(self, trainable, name):  # 设计策略网络
        data = keras.Input(shape=(self.s_dim,))
        x = keras.layers.Dense(30, activation="relu", trainable=trainable)(data)
        x = keras.layers.Dense(30, activation="relu", trainable=trainable)(x)
        x = keras.layers.Dense(self.a_dim, trainable=trainable)(x)
        a = self.a_bound * tf.math.tanh(x)
        model = keras.Model(data, a, name=name)
        return model

    def _build_critic(self, trainable, name):  # 设计评价网络
        data = keras.Input(shape=(self.a_dim + self.s_dim,))
        x = keras.layers.Dense(30, activation="relu", trainable=trainable)(data)
        x = keras.layers.Dense(30, activation="relu", trainable=trainable)(x)
        q = keras.layers.Dense(1, trainable=trainable)(x)
        model = keras.Model(data, q, name=name)
        return model

    def param_replace(self):  # 参数更新
        actor_weights = self.actor.get_weights()
        critic_weights = self.critic.get_weights()
        actor_target_weights = self.actor_.get_weights()
        critic_target_weights = self.critic_.get_weights()
        for i in range(len(actor_target_weights)):
            actor_target_weights[i] = actor_target_weights[i] * (1 - self.tau) + self.tau * actor_weights[i]
        for i in range(len(critic_target_weights)):
            critic_target_weights[i] = critic_target_weights[i] * (1 - self.tau) + self.tau * critic_weights[i]
        self.actor_.set_weights(actor_target_weights)
        self.critic_.set_weights(critic_target_weights)

    def act(self, s):  # 根据当前状态s执行动作
        a = self.actor.predict(np.reshape(s, (-1, self.s_dim)), verbose=0)[0]  # 使用策略网络
        return a

    def sample_memory(self):  # 从记忆库中采样数据
        indices = np.random.choice(self.memory_capacity, size=self.batch_size)
        bt = self.memory[indices, :]  # 获取批次数据
        bs = bt[:, :self.s_dim]
        ba = bt[:, self.s_dim: self.s_dim + self.a_dim]
        br = bt[:, -self.s_dim - 1: -self.s_dim]
        bs_ = bt[:, -self.s_dim:]
        return bs, ba, br, bs_

    def learn(self):  # 训练策略网络和Q网络
        bs, ba, br, bs_ = self.sample_memory()
        with tf.GradientTape() as tape:  更新策略网络
            a = self.actor(bs)  # 获取策略网络执行的动作
            q = self.critic(tf.concat([bs, a], 1))
            actor_loss = tf.reduce_mean(-q)  # 最大化价值函数值Q等于最小化-Q
        grads = tape.gradient(actor_loss, self.actor.trainable_variables)  # 仅更新策略网络参数
        self.a_opt.apply_gradients(zip(grads, self.actor.trainable_variables))

        with tf.GradientTape() as tape:  # 更新价值网络
            a_ = self.actor_(bs_)  # 目标策略网络根据下一状态决定下一动作
            q_ = br + self.gamma * self.critic_(tf.concat([bs_, a_], 1))
            q = self.critic(tf.concat([bs, ba], 1))
            critic_loss = self.mse(q_, q)  # 均方差损失函数
        grads = tape.gradient(critic_loss, self.critic.trainable_variables)  # 仅更新价值网络参数
        self.c_opt.apply_gradients(zip(grads, self.critic.trainable_variables))
        return actor_loss.numpy(), critic_loss.numpy()

    def store_transition(self, s, a, r, s_):  # 保存数据到记忆库中
        transition = np.hstack((s, a, [r], s_))
        index = self.pointer % self.memory_capacity
        self.memory[index, :] = transition
        self.pointer += 1

3. gym环境测试

import gym
import numpy as np
model = DDPG(a_dim=1, s_dim=3, a_bound=[-2, 2], batch_size=128, tau=0.01, gamma=0.9, memory_capacity=10000)

env = gym.make('Pendulum-v1')
RENDER = False  # 是否渲染环境
env = env.unwrapped  # 取消限制
max_ep_step = 200

var = 3
for ep in range(200):
    s = env.reset(seed=1)  # 设置随机种子
    ep_reward = 0
    for step in range(max_ep_step):
        if RENDER:
            env.render()
        # 增加探索时的噪音
        a = model.act(s)
        a = np.clip(np.random.normal(a, var), -2, 2)    # 为行动选择添加随机性进行探索，action超过[-2,2]时做截断处理
        s_, r, done, info = env.step(a)

        # 将当前的状态,行为,回报,下一个状态存储到记忆库中
        model.store_transition(s, a, r/10, s_)

        # 达到记忆库容量的最大值
        if model.pointer > 10000:
            var *= .9995  # 衰减动作随机性
            model.learn()  # 开始学习
            model.param_replace()  # 参数更新

        s = s_
        ep_reward += r
        if step == max_ep_step -1:
            print('Episode:', ep, ' Reward: %i' % int(ep_reward), 'Explore: %.2f' % var, )
            if ep_reward > -300:  # 达到回合最大值且回合回报值大于-300,渲染环境
                RENDER = True
            break

env.close()  # 关闭渲染窗口

基于tf1.x实现DDPG

1. 相关依赖库

• tensorflow==1.7.0
• gym==0.25.0
• python==3.6.0
• 测试运行环境——Windows

2. DDPG代码实现

import tensorflow as tf
import numpy as np

LR_A = 0.001  # 演员网络学习率
LR_C = 0.001  # 评论家网络学习率
GAMMA = 0.9  # 回报的折扣因子
TAU = 0.01  # 滑动平均数值
MEMORY_CAPACITY = 10000  # 记忆库大小
BATCH_SIZE = 32  # 训练批次

class DDPG(object):
    def __init__(self, a_dim, s_dim, a_bound):
        self.memory = np.zeros((MEMORY_CAPACITY, s_dim * 2 + a_dim + 1), dtype=np.float32)  # 设置记忆库存储的结构
        self.pointer = 0    # 记忆库当前容量
        self.sess = tf.Session()    # tf session会话
        # self.a_replace_counter, self.c_replace_counter = 0, 0   # 演员网络替换次数,评论家网络替换次数

        self.a_dim, self.s_dim, self.a_bound = a_dim, s_dim, a_bound,  # 动作空间,状态空间,动作范围
        self.S = tf.placeholder(tf.float32, [None, s_dim], 's')  # 当前状态预留占位符
        self.S_ = tf.placeholder(tf.float32, [None, s_dim], 's_')  # 下一状态预留占位符
        self.R = tf.placeholder(tf.float32, [None, 1], 'r')  # 回报预留占位符

        with tf.variable_scope('Actor'):
            self.a = self._build_a(self.S, scope='eval', trainable=True)  # 演员-评估网络根据当前状态输出行为
            a_ = self._build_a(self.S_, scope='target', trainable=False)   # 演员-目标网络根据下一个状态输出行为
        with tf.variable_scope('Critic'):
            # 当为td_error计算q时，在内存中分配self.a = a，否则当更新Actor时self.a来自Actor
            q = self._build_c(self.S, self.a, scope='eval', trainable=True)  # 根据当前状态和来自演员-评估网络的行为,计算q值
            q_ = self._build_c(self.S_, a_, scope='target', trainable=False)  # 根据下一状态和来自演员-目标网络的行为,计算q_值

        # 网络参数
        self.ae_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Actor/eval')
        self.at_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Actor/target')
        self.ce_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Critic/eval')
        self.ct_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Critic/target')

        # 目标网络参数替换(简单替换)
        self.soft_replace = [[tf.assign(ta, (1 - TAU) * ta + TAU * ea), tf.assign(tc, (1 - TAU) * tc + TAU * ec)]
                             for ta, ea, tc, ec in zip(self.at_params, self.ae_params, self.ct_params, self.ce_params)]

        q_target = self.R + GAMMA * q_
        # 在td_error的feed_dic中，self.a应该更改为内存中的行为
        td_error = tf.losses.mean_squared_error(labels=q_target, predictions=q)
        self.ctrain = tf.train.AdamOptimizer(LR_C).minimize(td_error, var_list=self.ce_params)  # 根据td_error进行更新评论家网络

        # 根据为代码进行修改: 相当于dq / da * da / dparams,即可以直接对q求梯度即可
        self.policy_grads = tf.gradients(ys=self.a, xs=self.ae_params, grad_ys=tf.gradients(q, self.a)[0])
        self.atrain = tf.train.AdamOptimizer(-LR_A).apply_gradients(zip(self.policy_grads, self.ae_params))

        self.sess.run(tf.global_variables_initializer())

    def choose_action(self, s):
        return self.sess.run(self.a, {self.S: s[np.newaxis, :]})[0]

    def learn(self):
        # 简单目标网络参数的替换
        self.sess.run(self.soft_replace)

        # 随机选择记忆库中BATCH_SIZE各数据进行更新
        indices = np.random.choice(MEMORY_CAPACITY, size=BATCH_SIZE)
        bt = self.memory[indices, :]
        bs = bt[:, :self.s_dim]
        ba = bt[:, self.s_dim: self.s_dim + self.a_dim]
        br = bt[:, -self.s_dim - 1: -self.s_dim]
        bs_ = bt[:, -self.s_dim:]

        # 演员网络更新
        self.sess.run(self.atrain, {self.S: bs})
        # 评论家网络更新
        self.sess.run(self.ctrain, {self.S: bs, self.a: ba, self.R: br, self.S_: bs_})

    def store_transition(self, s, a, r, s_):
        transition = np.hstack((s, a, [r], s_))
        index = self.pointer % MEMORY_CAPACITY  # 使用新的记忆替换旧的
        self.memory[index, :] = transition
        self.pointer += 1

    def _build_a(self, s, scope, trainable):
        with tf.variable_scope(scope):
            net = tf.layers.dense(s, 30, activation=tf.nn.relu, name='l1', trainable=trainable)
            a = tf.layers.dense(net, self.a_dim, activation=tf.nn.tanh, name='a', trainable=trainable)
            return tf.multiply(a, self.a_bound, name='scaled_a')

    def _build_c(self, s, a, scope, trainable):
        with tf.variable_scope(scope):
            n_l1 = 30
            w1_s = tf.get_variable('w1_s', [self.s_dim, n_l1], trainable=trainable)
            w1_a = tf.get_variable('w1_a', [self.a_dim, n_l1], trainable=trainable)
            b1 = tf.get_variable('b1', [1, n_l1], trainable=trainable)
            net = tf.nn.relu(tf.matmul(s, w1_s) + tf.matmul(a, w1_a) + b1)
            return tf.layers.dense(net, 1, trainable=trainable)  # Q(s,a)

3. gym环境测试

import numpy as np
import gym
from RL_brain import DDPG

MAX_EPISODES = 200  # 最大回合数
MAX_EP_STEPS = 200  # 每个回合最大步数
MEMORY_CAPACITY = 10000 # 记忆库容量

RENDER = False  # 是否渲染环境
ENV_NAME = 'Pendulum-v1'    # 环境名称

env = gym.make(ENV_NAME)    # 加载环境
env = env.unwrapped # 取消限制
env.seed(1) # 设置种子

s_dim = env.observation_space.shape[0]  # 状态空间
a_dim = env.action_space.shape[0]   # 行为空间
a_bound = env.action_space.high # 行为值上限

ddpg = DDPG(a_dim, s_dim, a_bound)  # 创建DDPG决策类

var = 3  # 控制探索
for i in range(MAX_EPISODES):
    s = env.reset()
    ep_reward = 0
    for j in range(MAX_EP_STEPS):
        if RENDER:
            env.render()

        # 增加探索时的噪音
        a = ddpg.choose_action(s)
        a = np.clip(np.random.normal(a, var), -2, 2)    # 为行动选择添加随机性进行探索
        s_, r, done, info, _ = env.step(a)

        # 将当前的状态,行为,回报,下一个状态存储到记忆库中
        ddpg.store_transition(s, a, r / 10, s_)

        # 达到记忆库容量的最大值
        if ddpg.pointer > MEMORY_CAPACITY:
            var *= .9995    # 衰减动作随机性
            ddpg.learn()    # 开始学习

        s = s_
        ep_reward += r
        if j == MAX_EP_STEPS-1:
            print('Episode:', i, ' Reward: %i' % int(ep_reward), 'Explore: %.2f' % var, )
            if ep_reward > -300:  # 达到回合最大值且回合回报值大于-300,渲染环境
                RENDER = True
            break

参考资料

• tf1.x代码实现 —— 书籍《强化学习》，邹伟，鬲玲，刘昱杓著，清华大学出版社；
• tf2.x代码实现 —— morvan强化学习机械臂，gitee链接：https://gitee.com/TIPE/robotArm