强化学习实践七 || N步DQN

原理

代码需要改动的地方：

params = SimpleNamespace(**{
    
    
    'env_name':         "CartPole-v0",
    'stop_reward':      250.0,
    'replay_size':      2000,       # buffer大小
    'replay_initial':   1000,       # buffer包含1000个数据时开始训练
    'target_net_sync':  300,        # 同步网络频率
    'epsilon_frames':   600,        # 分母
    'epsilon_start':    1.0,        # epsilon初始值
    'epsilon_final':    0.1,        # epsilon衰减后的稳定值
    'learning_rate':    0.001,
    'gamma':            0.90,       # 折扣回报率
    'batch_size':       32,
    'steps_count':      1           # 多步dqn的步长，若此参数为1，则为普通DQN
})

exp_source = ptan.experience.ExperienceSourceFirstLast(
    env, agent, gamma=params.gamma, steps_count=params.steps_count)

loss_v = common.calc_loss_dqn(buffer.sample(params.batch_size),
                                  net, tgt_net.target_model.to(device),
                                  params.gamma**params.steps_count,     # n步DQN，需要n次方折扣率
                                  device)

全部代码：

import gym
import ptan
import torch
import torch.optim as optim
import numpy as np
from lib import dqn_model,common
from tensorboardX import SummaryWriter
from types import SimpleNamespace

params = SimpleNamespace(**{
    
    
    'env_name':         "CartPole-v0",
    'stop_reward':      250.0,
    'replay_size':      2000,       # buffer大小
    'replay_initial':   1000,       # buffer包含1000个数据时开始训练
    'target_net_sync':  300,        # 同步网络频率
    'epsilon_frames':   600,        # 分母
    'epsilon_start':    1.0,        # epsilon初始值
    'epsilon_final':    0.1,        # epsilon衰减后的稳定值
    'learning_rate':    0.001,
    'gamma':            0.90,       # 折扣回报率
    'batch_size':       32,
    'steps_count':      1           # 多步dqn的步长，若此参数为1，则为普通DQN
})

writer = SummaryWriter(comment="_"+str(params.steps_count)+"_step_DQN")
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
env = gym.make(params.env_name)
env = env.unwrapped

STATE_SPACE = env.observation_space.shape[0]
ACTION_SPACE = env.action_space.n

net = dqn_model.DQN(STATE_SPACE, ACTION_SPACE).to(device)
tgt_net = ptan.agent.TargetNet(net)

selector = ptan.actions.ArgmaxActionSelector()
# epsilon_start表示随机动作的概率，该函数要与上面一行一起使用
selector = ptan.actions.EpsilonGreedyActionSelector(
    epsilon=params.epsilon_start, selector=selector)

epsilon_tracker = common.EpsilonTracker(selector, params)   # Epsilon衰减器

agent = ptan.agent.DQNAgent(net, selector, device=device)  # 定义agent

# 定义经验源和经验池
exp_source = ptan.experience.ExperienceSourceFirstLast(
    env, agent, gamma=params.gamma, steps_count=params.steps_count)
buffer = ptan.experience.ExperienceReplayBuffer(
    exp_source, buffer_size=params.replay_size)

# 定义优化器
optimizer = optim.Adam(net.parameters(), lr=params.learning_rate)

epsilon_step = 0
total_step = 0
step = 0
solved = False
mean_20reward = []
best_mean_reward = None
mean_reward = 0.0

while True:
    total_step += 1
    buffer.populate(1)      # 从经验源产生一个样本存入经验池，即执行一步动作

    # 每执行完一局就输出一些信息，如果没有执行完一局，就不会执行这个for循环
    for reward, game_steps in exp_source.pop_rewards_steps():
        step += 1
        print("第 %d 局游戏结束, reward=%.3f, epsilon=%.2f 共执行 %d 步 总步数 %d" % (
            step, reward, epsilon_tracker.selector.epsilon, game_steps, total_step))
        mean_20reward.append(reward)
        mean_reward = np.mean(mean_20reward[-20:])
        writer.add_scalar("mean_20reward", mean_reward, step)
        if best_mean_reward is None or mean_reward > best_mean_reward:
            if best_mean_reward is not None:
                print("最近20局平均奖励更新 %.2f => %.2f"%(best_mean_reward,mean_reward))
            best_mean_reward = mean_reward

        solved = mean_reward > params.stop_reward

    if solved:
        print("Congrats!")
        break

    if len(buffer) < params.replay_initial:
        continue

    # 训练
    optimizer.zero_grad()
    loss_v = common.calc_loss_dqn(buffer.sample(params.batch_size),
                                  net, tgt_net.target_model.to(device),
                                  params.gamma**params.steps_count,
                                  device)
    loss_v.backward()
    optimizer.step()

    writer.add_scalar("loss", loss_v.item(), epsilon_step)

    epsilon_step += 1
    # 随机选择动作的概率以每步乘以EPS_DECAY的频率衰减
    epsilon_tracker.frame(epsilon_step)
    writer.add_scalar("epsilon", epsilon_tracker.selector.epsilon, epsilon_step)


    # 同步更新训练网络的参数至目标网络
    if epsilon_step % params.target_net_sync == 0:
        tgt_net.sync()

writer.close()

比较：

综合来看，2步DQN效果好于普通DQN和3步DQN