強化學習經典算法筆記(八)：LSTM加持的A2C算法解決POMDP問題

強化學習經典算法筆記(八)：LSTM加持的A2C算法解決POMDP問題

最近用到LSTM構建Agent，找到了一個非常簡明易讀的示例代碼。
https://github.com/HaiyinPiao/pytorch-a2clstm-DRQN

環境採用CartPole-v1。原本狀態是一個4維向量，現刪去第二維，即小車的速度，保留小車的位移，杆的角度和角速度，使問題從MDP問題變爲POMDP（Partial Observable Markov Decision Process）問題。

代碼如下：

導入必要package

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import matplotlib.pyplot as plt
import numpy as np
import math
import random
import os
import gym

參數設置

STATE_DIM = 4-1;    # 刪去小車速度這一維度，使之成爲POMDP
ACTION_DIM = 2;     # 動作空間大小
NUM_EPISODE = 5000; # 訓練的Episode數量
EPISODE_LEN = 1000; # episode最大長度
A_HIDDEN = 40;      # Actor網絡的隱層神經元數量
C_HIDDEN = 40;      # Critic網絡的隱層神經元數量

ActorCritic網絡

# ActorNet使用LSTM + MLP估計完整的狀態
class ActorNetwork(nn.Module):

    def __init__(self,in_size,hidden_size,out_size):
        super(ActorNetwork, self).__init__()
        self.lstm = nn.LSTM(in_size, hidden_size, batch_first = True)
        self.fc = nn.Linear(hidden_size,out_size)

    def forward(self, x, hidden):
        x, hidden = self.lstm(x, hidden)
        x = self.fc(x)
        x = F.log_softmax(x,2)  # log(softmax(x))
        return x, hidden

class ValueNetwork(nn.Module):

    def __init__(self,in_size,hidden_size,out_size):
        super(ValueNetwork, self).__init__()
        self.lstm = nn.LSTM(in_size, hidden_size, batch_first = True)
        self.fc = nn.Linear(hidden_size,out_size)

    def forward(self,x, hidden):
        x, hidden = self.lstm(x, hidden)
        x = self.fc(x)
        return x, hidden

完成單episode交互並記錄trajectory

def roll_out(actor_network,env,episode_len,value_network,init_state):
    '''
    rollout最長1000frames
    返回：
    狀態序列，不包括終態
    動作序列，獨熱編碼
    獎勵序列，不包括終態獎勵
    state：遊戲環境初始化後的初始狀態
    '''
    states = []
    actions = []
    rewards = []
    is_done = False
    final_r = 0
    state = init_state # 初始狀態
    a_hx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0); # 初始化隱狀態
    a_cx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0);
    c_hx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);
    c_cx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);

    for j in range(episode_len):
        states.append(state)
        log_softmax_action, (a_hx,a_cx) = actor_network(Variable(torch.Tensor([state]).unsqueeze(0)), (a_hx,a_cx))
        # 這個部分可以用torch Categorical來實現
        # from torch.distributions import Categorical
        softmax_action = torch.exp(log_softmax_action) # 對數softmax取指數，保證大於0
        action = np.random.choice(ACTION_DIM,p=softmax_action.cpu().data.numpy()[0][0])
        
        # 動作獨熱編碼
        one_hot_action = [int(k == action) for k in range(ACTION_DIM)]
        
        next_state,reward,done,_ = env.step(action)
        next_state = np.delete(next_state, 1)
        #fix_reward = -10 if done else 1
        
        actions.append(one_hot_action)
        rewards.append(reward)
        final_state = next_state # final_state和state是一回事
        state = next_state
        if done:
            is_done = True
            state = env.reset()
            state = np.delete(state,1)
            a_hx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0);
            a_cx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0);
            c_hx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);
            c_cx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);

            # 打印episode總分
            print(j+1)
            break
    if not is_done: # 1000frame後如果episode還未結束，就用VNet估計終態價值c_out
        c_out, (c_hx,c_cx) = value_network(Variable(torch.Tensor([final_state])), (c_hx,c_cx))
        final_r = c_out.cpu().data.numpy() # 如果episode正常結束，final_r=0表示終態cart失去控制得0分
    return states,actions,rewards,final_r,state

計算累計折扣獎勵的函數

def discount_reward(r, gamma,final_r):
    '''
    r:          list
    final_r:    scalar
    '''
    discounted_r = np.zeros_like(r)
    running_add = final_r
    for t in reversed(range(0, len(r))):
        running_add = running_add * gamma + r[t]
        discounted_r[t] = running_add
    return discounted_r

訓練和測試主函數

def main():
    # 初始化env
    env = gym.make("CartPole-v1")
    init_state = env.reset()
    init_state = np.delete(init_state,1) # 刪掉cart velocity這一維度
    

    # 初始化價值網絡
    value_network = ValueNetwork(in_size=STATE_DIM, hidden_size=C_HIDDEN, out_size=1)
    value_network_optim = torch.optim.Adam(value_network.parameters(),lr=0.005)

    # 初始化動作網絡
    actor_network = ActorNetwork(in_size=STATE_DIM, hidden_size=A_HIDDEN, out_size=ACTION_DIM)
    actor_network_optim = torch.optim.Adam(actor_network.parameters(),lr = 0.001)

    steps =[]
    task_episodes =[]
    test_results =[]

    for episode in range(NUM_EPISODE):
        # 完成一輪rollout
        states,actions,rewards,final_r,current_state = roll_out(actor_network,env,EPISODE_LEN,value_network,init_state)
        # states.shape = [epi_len,3],list
        
        # rollout結束後的初態
        init_state = current_state
        actions_var = Variable(torch.Tensor(actions).view(-1,ACTION_DIM)).unsqueeze(0)
        states_var = Variable(torch.Tensor(states).view(-1,STATE_DIM)).unsqueeze(0)

        # 訓練動作網絡
        a_hx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0);
        a_cx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0);
        c_hx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);
        c_cx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);
        
        actor_network_optim.zero_grad()
        # print(states_var.unsqueeze(0).size())
        log_softmax_actions, (a_hx,a_cx) = actor_network(states_var, (a_hx,a_cx))
        vs, (c_hx,c_cx) = value_network(states_var, (c_hx,c_cx)) # 給出狀態價值估計
        vs.detach()    # 不參與求梯度
        
        # 計算Q(s,a)和Advantage函數
        qs = Variable(torch.Tensor(discount_reward(rewards,0.99,final_r)))
        qs = qs.view(1, -1, 1)
        advantages = qs - vs
        # print('adv,',advantages.shape)
        # log_softmax_actions * actions_var是利用獨熱編碼特性取出對應action的對數概率
        actor_network_loss = - torch.mean(torch.sum(log_softmax_actions*actions_var,1)* advantages)
        actor_network_loss.backward()
        torch.nn.utils.clip_grad_norm(actor_network.parameters(),0.5)
        actor_network_optim.step()

        # 訓練價值網絡
        value_network_optim.zero_grad()
        target_values = qs
        a_hx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0);
        a_cx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0);
        c_hx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);
        c_cx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);
        values, (c_hx,c_cx) = value_network(states_var, (c_hx,c_cx))

        criterion = nn.MSELoss()
        value_network_loss = criterion(values,target_values)
        value_network_loss.backward()
        torch.nn.utils.clip_grad_norm(value_network.parameters(),0.5)
        value_network_optim.step()

        # Testing
        if (episode + 1) % 50== 0:
                result = 0
                test_task = gym.make("CartPole-v1")
                for test_epi in range(10):       # 測試10個episode
                    state = test_task.reset()
                    state = np.delete(state,1)
                    
                    a_hx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0);
                    a_cx = torch.zeros(A_HIDDEN).unsqueeze(0).unsqueeze(0);
                    c_hx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);
                    c_cx = torch.zeros(C_HIDDEN).unsqueeze(0).unsqueeze(0);
                    
                    for test_step in range(500): # 每個episode最長500frame
                        
                        log_softmax_actions, (a_hx,a_cx) = actor_network(Variable(torch.Tensor([state]).view(1,1,3)), (a_hx,a_cx))
                        softmax_action = torch.exp(log_softmax_actions)
                        
                        #print(softmax_action.data)
                        action = np.argmax(softmax_action.data.numpy()[0])
                        next_state,reward,done,_ = test_task.step(action)
                        next_state = np.delete(next_state,1)
                        
                        result += reward
                        state = next_state
                        if done:
                            break
                print("episode:",episode+1,"test result:",result/10.0)
                steps.append(episode+1)
                test_results.append(result/10)
    plt.plot(steps,test_results)
    plt.savefig('training_score.png')
if __name__ == '__main__':
    main()

實驗結果如下：