add Qlearning

codes/QLearning/README.md

@@ -0,0 +1,19 @@
+## CliffWalking-v0环境简介
+
+悬崖寻路问题（CliffWalking）是指在一个4 x 12的网格中，智能体以网格的左下角位置为起点，以网格的下角位置为终点，目标是移动智能体到达终点位置，智能体每次可以在上、下、左、右这4个方向中移动一步，每移动一步会得到-1单位的奖励。
+
+<img src="assets/image-20201007211441036.png" alt="image-20201007211441036" style="zoom:50%;" />
+
+如图，红色部分表示悬崖，数字代表智能体能够观测到的位置信息，即observation，总共会有0-47等48个不同的值，智能体再移动中会有以下限制：
+
+* 智能体不能移出网格，如果智能体想执行某个动作移出网格，那么这一步智能体不会移动，但是这个操作依然会得到-1单位的奖励
+
+* 如果智能体“掉入悬崖” ，会立即回到起点位置，并得到-100单位的奖励
+
+* 当智能体移动到终点时，该回合结束，该回合总奖励为各步奖励之和
+
+实际的仿真界面如下：
+
+<img src="assets/image-20201007211858925.png" alt="image-20201007211858925" style="zoom:50%;" />
+
+由于从起点到终点最少需要13步，每步得到-1的reward，因此最佳训练算法下，每个episode下reward总和应该为-13。

codes/QLearning/agent.py

@@ -0,0 +1,63 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-09-11 23:03:00
+LastEditor: John
+LastEditTime: 2021-03-11 19:16:27
+Discription: 
+Environment: 
+'''
+from functools import update_wrapper
+import numpy as np
+import math
+import torch
+from collections import defaultdict
+
+class QLearning(object):
+    def __init__(self,
+                 n_actions,cfg):
+        self.n_actions = n_actions  # number of actions
+        self.lr = cfg.lr  # learning rate
+        self.gamma = cfg.gamma  
+        self.epsilon = 0 
+        self.sample_count = 0  # epsilon随训练的也就是采样次数逐渐衰减，所以需要计数
+        self.epsilon_start = cfg.epsilon_start
+        self.epsilon_end = cfg.epsilon_end
+        self.epsilon_decay = cfg.epsilon_decay
+        self.Q_table  = defaultdict(lambda: np.zeros(n_actions)) # 使用字典存储Q表，个人比较喜欢这种，也可以用下面一行的二维数组表示，但是需要额外更改代码
+        # self.Q_table = np.zeros((n_states, n_actions))  # Q表
+    def choose_action(self, state):
+        self.sample_count += 1
+        self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
+            math.exp(-1. * self.sample_count / self.epsilon_decay)
+        # 随机选取0-1之间的值，如果大于epsilon就按照贪心策略选取action，否则随机选取
+        if np.random.uniform(0, 1) > self.epsilon:
+            action = np.argmax(self.Q_table[state])
+        else:
+            action = np.random.choice(self.n_actions)  # 有一定概率随机探索选取一个动作
+        return action
+            
+    def update(self, state, action, reward, next_state, done):
+        Q_predict = self.Q_table[state][action]
+        if done:
+            Q_target = reward  # terminal state
+        else:
+            Q_target = reward + self.gamma * np.max(
+                self.Q_table[next_state])  # Q_table-learning
+        self.Q_table[state][action] += self.lr * (Q_target - Q_predict)
+    def save(self,path):
+        '''把 Q表格 的数据保存到文件中
+        '''
+        import dill
+        torch.save(
+            obj=self.Q_table,
+            f=path,
+            pickle_module=dill
+        )
+
+    def load(self, path):
+        '''从文件中读取数据到 Q表格
+        '''
+        self.Q_table =torch.load(f='prod_dls.pkl',pickle_module=dill)

codes/QLearning/main.py

@@ -0,0 +1,115 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-09-11 23:03:00
+LastEditor: John
+LastEditTime: 2021-03-11 19:22:50
+Discription: 
+Environment: 
+'''
+
+import sys,os
+sys.path.append(os.getcwd()) # 添加当前终端路径
+import argparse
+import gym
+import datetime
+from QLearning.plot import plot
+from QLearning.utils import save_results
+from envs.gridworld_env import CliffWalkingWapper, FrozenLakeWapper
+from QLearning.agent import QLearning
+
+SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
+SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/'
+RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/result/"+SEQUENCE+'/'
+
+def get_args():
+    '''训练的模型参数
+    '''
+    parser = argparse.ArgumentParser()
+    '''训练相关参数'''
+    parser.add_argument("--n_episodes", default=500,
+                        type=int, help="训练的最大episode数目")       
+    '''算法相关参数'''
+    parser.add_argument("--gamma", default=0.9,
+                        type=float, help="reward的衰减率")
+    parser.add_argument("--epsilon_start", default=0.99,
+                        type=float, help="e-greedy策略中初始epsilon")
+    parser.add_argument("--epsilon_end", default=0.01,
+                        type=float, help="e-greedy策略中的结束epsilon")
+    parser.add_argument("--epsilon_decay", default=200,
+                        type=float, help="e-greedy策略中epsilon的衰减率")
+    parser.add_argument("--lr", default=0.1, type=float, help="学习率")
+    config = parser.parse_args()
+    return config
+def train(cfg,env,agent):
+    # env = gym.make("FrozenLake-v0", is_slippery=False)  # 0 left, 1 down, 2 right, 3 up
+    # env = FrozenLakeWapper(env)
+    rewards = []  # 记录所有episode的reward,
+    steps = []  # 记录所有episode的steps
+    for i_episode in range(cfg.n_episodes):
+        ep_reward = 0  # 记录每个episode的reward
+        ep_steps = 0  # 记录每个episode走了多少step
+        obs = env.reset()  # 重置环境, 重新开一局（即开始新的一个episode）
+        while True:
+            action = agent.choose_action(obs)  # 根据算法选择一个动作
+            next_obs, reward, done, _ = env.step(action)  # 与环境进行一个交互
+            # 训练 Q-learning算法
+            agent.update(obs, action, reward, next_obs, done)  # 不需要下一步的action
+            obs = next_obs  # 存储上一个观察值
+            ep_reward += reward
+            ep_steps += 1  # 计算step数
+            if done:
+                break
+        steps.append(ep_steps)
+        # 计算滑动平均的reward
+        if rewards:
+            rewards.append(rewards[-1]*0.9+ep_reward*0.1)
+        else:
+            rewards.append(ep_reward)
+        print("Episode:{}/{}: reward:{:.1f}".format(i_episode+1, cfg.n_episodes,ep_reward))
+    plot(rewards)
+    if not os.path.exists(SAVED_MODEL_PATH):
+        os.mkdir(SAVED_MODEL_PATH)
+    agent.save(SAVED_MODEL_PATH+'Q_table.pkl')  # 训练结束，保存模型
+    '''存储reward等相关结果'''
+    save_results(rewards,tag='train',result_path=RESULT_PATH)
+
+def eval(cfg,env,agent):
+    # env = gym.make("FrozenLake-v0", is_slippery=False)  # 0 left, 1 down, 2 right, 3 up
+    # env = FrozenLakeWapper(env)
+    rewards = []  # 记录所有episode的reward,
+    steps = []  # 记录所有episode的steps
+    for i_episode in range(20):
+        ep_reward = 0  # 记录每个episode的reward
+        ep_steps = 0  # 记录每个episode走了多少step
+        obs = env.reset()  # 重置环境, 重新开一局（即开始新的一个episode）
+        while True:
+            action = agent.choose_action(obs)  # 根据算法选择一个动作
+            next_obs, reward, done, _ = env.step(action)  # 与环境进行一个交互
+            obs = next_obs  # 存储上一个观察值
+            ep_reward += reward
+            ep_steps += 1  # 计算step数
+            if done:
+                break
+        steps.append(ep_steps)
+        # 计算滑动平均的reward
+        if rewards:
+            rewards.append(rewards[-1]*0.9+ep_reward*0.1)
+        else:
+            rewards.append(ep_reward)
+        print("Episode:{}/{}: reward:{:.1f}".format(i_episode+1, cfg.n_episodes,ep_reward))
+    plot(rewards)
+    '''存储reward等相关结果'''
+    save_results(rewards,tag='eval',result_path=RESULT_PATH)
+    
+if __name__ == "__main__":
+    cfg = get_args()
+    env = gym.make("CliffWalking-v0")  # 0 up, 1 right, 2 down, 3 left
+    env = CliffWalkingWapper(env)
+    n_actions = env.action_space.n
+    agent = QLearning(n_actions,cfg)
+    train(cfg,env,agent)
+    eval(cfg,env,agent)
+    

codes/QLearning/plot.py

@@ -0,0 +1,35 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-10-07 20:57:11
+LastEditor: John
+LastEditTime: 2020-10-07 21:00:29
+Discription: 
+Environment: 
+'''
+import matplotlib.pyplot as plt
+import seaborn as sns
+import numpy as np
+import os 
+
+def plot(item,ylabel='rewards'):
+    sns.set()
+    plt.figure()
+    plt.plot(np.arange(len(item)), item)
+    plt.title(ylabel+' of Q-learning') 
+    plt.ylabel(ylabel)
+    plt.xlabel('episodes')
+    plt.savefig(os.path.dirname(__file__)+"/result/"+ylabel+".png")
+    plt.show()
+
+if __name__ == "__main__":
+
+    output_path = os.path.dirname(__file__)+"/result/"
+    rewards=np.load(output_path+"rewards_train.npy", )
+    MA_rewards=np.load(output_path+"MA_rewards_train.npy")
+    steps = np.load(output_path+"steps_train.npy")
+    plot(rewards)
+    plot(MA_rewards,ylabel='moving_average_rewards')
+    plot(steps,ylabel='steps')

codes/QLearning/utils.py

@@ -0,0 +1,22 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-11-23 13:44:52
+LastEditor: John
+LastEditTime: 2021-03-11 19:18:34
+Discription: 
+Environment: 
+'''
+import os
+import numpy as np
+
+
+def save_results(rewards,tag='train',result_path='./result'):
+    '''保存reward等结果
+    '''
+    if not os.path.exists(result_path): # 检测是否存在文件夹
+        os.mkdir(result_path)
+    np.save(result_path+'rewards_'+tag+'.npy', rewards)
+    print('results saved!')