这是一个易理解的 demo，300行左右，可以作为RL的入门代码，辅助基础公式的理解

这个是我自己的学习笔记。三连留下邮箱，可以直接发送完整的代码标注文件~

如有错误，麻烦指出！我已经蛮久没写博了，上一篇RL博客也快一年半了，很久没做这一块了。硕士刚入学，兜兜转转还是回到了RL。欢迎交流~

井字棋规则：感觉就是三子棋。3 * 3棋盘，先连成3子胜利。

代码概览

🌱 类的定义：定义了State、Judger、Player、HumanPlayer四个类，分别代表棋局状态、下棋(裁判)、AI棋手、人类棋手

🌱 状态s：每个棋局是一种状态，使用hash标识唯一的状态。共3^9种状态

🌱 训练过程：首先让2个AI棋手对战，以逐渐完善策略(价值状态函数)。AI棋手训练完后，让AI棋手和人类棋手对战

🌱 训练AI棋手时：初始时，设置胜局状态value为1，输局状态value为0，其余为0.5。然后backup更新，即利用 V(s) ⬅ V(s) + α[V(s')-V(s)] 不断修正value，直到逐渐收敛。α是步长

代码讲解

main主函数

先来看下主函数：

if __name__ == '__main__':train(int(1e5))   # 1e5是浮点型。是epoch  # 2AI对战，完善value functioncompete(int(1e3))   # 2AI对战训练完后，再对战自测胜率 play()   # 人类和AI对战

总之就是先AI对战，再人机大战

State状态类

引入包并定义3*3棋局：

import numpy as np
import pickleBOARD_ROWS = 3
BOARD_COLS = 3
BOARD_SIZE = BOARD_ROWS * BOARD_COLS

再进行State状态类的定义。State类包括函数：

🌻 __init__()初始化、

🌻 hash()计算每个状态的哈希值以索引、

🌻 is_end()检查棋局是否结束、

🌻 next_station()函数将棋手标志放至下一个下棋位置上、

🌻 print()打印当前3*3棋局 

详细注释见代码：

# 每一个state是棋盘的整个状态，共3^9个状态
class State:def __init__(self):# 1 symbol： 先行player# -1 symbol：后行player# 0 symbol：empty positionself.data = np.zeros((BOARD_ROWS, BOARD_COLS))  # 代表boardself.winner = Noneself.hash_val = None    # 使用hash标识每个状态self.end = None# 计算每个状态的哈希值（规则随机）def hash(self):if self.hash_val is None:self.hash_val = 0for i in self.data.reshape(BOARD_COLS * BOARD_ROWS):if i == -1:i = 2self.hash_val = self.hash_val * 3 + ireturn int(self.hash_val)# 检查游戏是否分出胜负，或是平局def is_end(self):if self.end is not None:return self.endresults = []# check rowfor i in range(0, BOARD_ROWS):results.append(np.sum(self.data[i, :]))# check columnsfor i in range(0, BOARD_COLS):results.append(np.sum(self.data[i, :]))# check diagnoalsresults.append(0)for i in range(0, BOARD_ROWS):results[-1] += self.data[i, i]results.append(0)for i in range(0, BOARD_COLS):results[-1] += self.data[i, BOARD_ROWS -1 - i]for result in results:if result == 3:self.end = Trueself.winner = 1return self.endif result == -3:self.end = Trueself.winner = -1return self.end# check tiesum = np.sum(np.abs(self.data))if sum == BOARD_COLS * BOARD_ROWS:self.end = Trueself.winner = 0  # 平局return self.end# 非胜负/平局，继续游戏self.end = Falsereturn self.end# 下一个状态# 将棋手标志放置board位置(i, j)def next_station(self, i, j, symbol):new_state = State()new_state.data = np.copy(self.data)new_state.data[i, j] = symbolreturn new_state# 打印棋局def print(self):for i in range(0, BOARD_ROWS):print('-----------------')out = '| 'for j in range(0, BOARD_COLS):if self.data[i, j] == 1:token = '*'if self.data[i, j] == -1:token = 'x'if self.data[i, j] == 0:token = '0'out += token + ' | 'print(out)print('-----------------')

# 检索当前状态下，所有下一个可能动作带来的状态变换
def get_all_states_impl(current_state, current_symbol, all_states):for i in range(0, BOARD_ROWS):for j in range(0, BOARD_COLS):if current_state.data[i][j] == 0:   # 检索目前所有空格子newState = current_state.next_state(i, j, current_symbol)newHash = newState.hash()if newHash not in all_states.keys():isEnd = newState.is_end()all_states[newHash] = (newState, isEnd)if not isEnd:   # 如果棋手1下完还没结束棋局，棋手2下get_all_states_impl(newState, -current_symbol, all_states)def get_all_states():current_symbol = 1current_state = State()all_states = dict()all_states[current_state.hash()] = (current_state, current_state.is_end())get_all_states_impl(current_state, current_symbol, all_states)return all_states# all_states字典：key是某状态对应的唯一哈希值，value是(state,isEnd)
all_states = get_all_states()

Judger裁判类

Judger类是裁判，其实就是两个棋手轮流下棋。包括函数：

🌻 __init__() 初始化

🌻 reset() 重置

🌻 alternate() 轮流选择下棋手

🌻 play() 双方下棋

详细注释见代码：

class Judger:def __init__(self, player1, player2):self.p1 = player1self.p2 = player2self.current_player = Noneself.p1_symbol = 1self.p2_symbol = -1self.p1.set_symbol(self.p1_symbol)self.p2.set_symbol(self.p2_symbol)def reset(self):self.p1.reset()self.p2.reset()def alternate(self):while True:yield self.p1yield self.p2# play函数用于双方轮流下棋（这个play函数是两个AI棋手下），# act函数用于为当前player选择value最高的下棋位置# @print:if True, print each board during the gamedef play(self, print = False):alternator = self.alternate()self.reset()current_state = State()self.p1.set_state(current_state)self.p1.set_state(current_state)while True: # 一直到棋局结束，return了才结束循环player = next(alternator)   # 双方轮流下棋if print:current_state.print()[i, j, symbol] = player.act()    # 为棋手选择下一步最佳落子next_state_hash = current_state.next_state(i, j, symbol)current_state, is_end = all_states[next_state_hash]self.p1.set_state(current_state)self.p2.set_state(current_state)if is_end:if print:current_state.print()return current_state.winner

Player棋手(AI)类

Player是AI棋手类。包括函数：

🌻 __init__() 初始化

🌻 reset() 重置

🌻 set_state() 设置状态及是否explore

🌻 set_symbol() 状态价值初始化赋值

🌻 backup() 反向更新状态价值：V(s) ⬅ V(s) + α[V(s')-V(s)]

🌻 act() 当前state下，选择最优action

🌻 save_policy() 保存策略(就是estimations价值)

🌻 load_policy() 加载策略

详细注释见代码：

# AI player
# 关于value function(state的函数)解释：https://face2ai.com/RL-RSAB-1-5-An-Extended-Example/
class Player:# @step_size: the step size to update estimation# (好像就是value function)，back up更新里的α，详见上方链接解释# @epsilon: the probability to exploredef __init__(self, step_size = 0.1, epsilon = 0.1):self.estimations = dict()self.step_size = step_sizeself.epsilon = epsilonself.states = []self.greedy = []def reset(self):self.states = []self.greedy = []def set_state(self, state):self.states.append(state)self.greedy.append(True)    # 应该是exploit而完全不explore# 这个函数就是给状态state赋value的(estimation字典，key为hash(对应某个状态)，值为value)，# 在棋局结束状态下，如果这个状态赢，赋1；若平局，赋0.5，输则赋0# 正在进行时的棋局状态一律赋0.5def set_symbol(self, symbol):self.symbol = symbolfor hash_val in all_states.keys():(state, is_end) = all_states[hash_val]if is_end:if state.winner == self.symbol:self.estimations[hash_val] = 1.0elif state.winner == 0:self.estimations[hash_val] = 0.5else:self.estimations[hash_val] = 0else:self.estimations[hash_val] = 0.5# 反向更新价值状态函数（value estimation）def backup(self):self.states = [state.hash() for state in self.states]# 反向更新：V(s) ⬅ V(s) + α[V(s')-V(s)]# 可参考链接：https://face2ai.com/RL-RSAB-1-5-An-Extended-Example/for i in reversed(range(len(self.states) - 1)):state = self.states[i]td_error = self.greedy[i] * (self.estimations[self.states[i + 1]] - self.estimations[state])self.estimations[state] += self.step_size * td_error# 当前state下，选择下一步的最优action# act函数的返回结果为下棋位置和棋手标志(i, j, symbol)def act(self):state = self.states[-1]next_states = []next_positions = []# 找出目前state下所有空位for i in range(BOARD_ROWS):for j in range(BOARD_COLS):if state.data[i, j] == 0:next_positions.append([i,j])next_states.append(state.next_state(i, j, self.symbol).hash())# exploreif np.random.rand() < self.epsilon:    # 随机生成(0,1)之间数 action = next_positions[np.random.randint(len(next_positions))]action.append(self.symbol)self.greedy[-1] = Falsereturn action# 否则exploitvalues = []for hash, pos in zip(next_states, next_positions):values.append((self.estimations[hash], pos))np.random.shuffle(values)values.sort(key = lambda x:x[0], reverse = True)    # 按照state的value值大小倒序排action = values[0][1]    # 选择value值最大的action的位置action.append(self.symbol)  # 为这个动作加上棋手标志return actiondef save_policy(self):# bin是二进制格式的文件with open('policy_%s.bin' % ('first' if self.symbol == 1 else 'second'), 'wb') as f:pickle.dump(self.estimations, f)    # 对象存储def load_policy(self):with open('policy_%s.bin' % ('first' if self.symbol == 1 else 'second'), 'rb') as f:self.estimations = pickle.load(f)

HumanPlayer棋手(人类)类

HumanPlayer类就是人类棋手。包括函数：

🌻 __init__() 初始化

🌻 set_state() 设置状态

🌻 set_symbol() 设置棋手标志

🌻 act() 人类棋手通过键盘下棋

详细注释见代码：

# human interface
# input a number to put a chessman
# | q | w | e |
# | a | s | d |
# | z | x | c |
class HumanPlayer:def __init__(self, **kwargs):self.symbol = Noneself.keys = ['q', 'w', 'e', 'a', 's', 'd', 'z', 'x', 'c']self.state = Nonereturndef reset(self):returndef set_state(self, state):self.state = statedef set_symbol(self, symbol):self.symbol = symbolreturndef backup(self, _):returndef act(self):self.state.print()key = input("Input your position:")  # 将这句话显示在屏幕上，接收用户输入的值，赋给key# 默认用户的输入是键盘最左边三行三列字母data = self.keys.index(key)  # 索引i = data // int(BOARD_COLS)j = data % BOARD_COLSreturn (i, j, self.symbol)

train函数

train()其实就是让两个AI对战，不断完善value function，直至逐渐收敛。是在训练AI棋手

# 2个AI间训练
# 两个AI player打，不断完善value function(即estimations)
# 训练结束后, Epoch 10000, player 1 win 0.08, player 2 win 0.03
def train(epochs):player1 = Player(epsilon = 0.01)player2 = Player(epsilon = 0.01)judger = Judger(player1, player2)player1_win = 0.0player2_win = 0.0for i in range(1, epochs + 1):winner = judger.play(print = False)   # 开始下棋if winner == 1:player1_win += 1if winner == -1:player2_win += 1print('Epoch %d, player1 win %.02f, player2 win %.02f' % (i, player1_win / i, player2_win / i))player1.backup()player2.backup()judger.reset()player1.save_policy()player2.save_policy()

compete函数

相当于train后的test。这个游戏规则太简单了，测试时两个AI都是平局

# 2个AI间测试
# 训练结束后的测试，AI之间就没有输赢了，全是平局：
# 1000 turns, player 1 win 0.00, player 2 win 0.00
# 计算下turns次棋，两个AI棋手的分别胜率
def compete(turns):player1 = Player(epsilon = 0)player2 = Player(epsilon = 0)judger = Judger(player1, player2)player1.load_policy()player2.load_policy()player1_win = 0.0player2_win = 0.0for i in range(0, turns):winner  = judger.play()if winner == 1:player1_vin += 1if winner == -1:player2_win += 1judger.reset()print('%d turns, player1 win %.02f, player1 win %.02f' % (turns, player1_win / turns, player2_win / turns))

play函数

人类和AI下棋

# 人类和AI下棋
def play():while True:player1 = HumanPlayer()player2 = Player(epsilon = 0)judger = Judger(player1, player2)player2.load_policy()   # AI棋手使用之前两个AI对战储存的policy（即value function，因为上面设置的epsilon为0，直接贪婪地选造成value最大状态的action）winner = judger.play()if winner == player2.symbol:print("You lose!")elif winner == player1.symbol:print("You win!")else:print("It is a tie!")