一、游戏环境：

冰面求生

玩家从左上角绿色位置跑到右下角黄色位置，每次只能选择上下左右走一步，深绿色是安全区，紫色的位置是陷阱，掉下去就没命。玩家是不知道陷阱在哪的，他需要不断的从失败中学习，获得知识，最后轻盈的走到黄色奖品区。

环境代码：

import numpy as np
from matplotlib import pyplot as plt
import random
from itertools import count

class Env:
    def __init__(self):
        self.action_space = []
        self.agent = None
        self.env = np.zeros((4,12))
        self.env[-1,1:-1] = -1
        self.env[-1,-1] = 1
        self.env_show = self.env.copy()

    def step(self,a=[0,0],ai=False):
        s_ = np.array(self.agent.s) + np.array(a)
        if 0<=s_[0]<=3 and 0<=s_[1]<=11:
            self.agent.s = s_
            r = self.env[s_[0],s_[1]]
        else:
            s_ = self.agent.s
            r = -1
        self.agent.s_chain = np.vstack([self.agent.s_chain,s_])
        self.agent.sar_chain = np.vstack([self.agent.sar_chain,np.hstack([s_,a,r])])
        return s_,r
    
    def play(self):
        env.reset()
        for t in count():
            a = agent.chose_action()
            s,r = env.step(a)
            if r in [-1,1]:
                break
        env.render()
        return t,r

    def render(self):
        for i,j in self.agent.s_chain:
            self.env_show[i,j] = 0.5
        plt.imshow(self.env_show)

    def reset(self):
        self.agent.reset()
        self.env_show = self.env.copy()

    def register(self,agent):
        self.agent = agent

环境中必备step函数，输入每次动作返回新位置与奖励值，这里生存奖励0，角色失败返回-1.

二、随机到处乱跑的人工智能

class Agent:
    def __init__(self):
        self.action_space = np.array([[0,1],[-1,0],[0,-1],[1,0]])
        self.s = np.array([0,0])
        self.s_chain = np.expand_dims(self.s,0)
        self.sar_chain = np.expand_dims(np.hstack([np.array([0,0]),[0,0],0]),0)

    def chose_action(self,ai=False):
        a = random.choice(self.action_space)
        return a

    def learn(self):
        ...

    def reset(self):
        self.action_space = np.array([[0,1],[-1,0],[0,-1],[1,0]])
        self.s = np.array([0,0])
        self.s_chain = np.expand_dims(self.s,0)
        self.sar_chain = np.expand_dims(np.hstack([np.array([0,0]),[0,0],0]),0)

env = Env()
agent = Agent()
env.register(agent)

三、开始游戏

给环境添加 play_until_success 函数：

def play_until_success(self):
        for t in count():
            _,r = self.play()
            if t%20000 == 0:
                print(f"playing {t} times!")
            if r == 1:
                print(f"{t} times success!")
                self.render()
                break

run一下：

playing 0 times!
playing 20000 times!
playing 40000 times!
42777 times success!

success！

经过 4.2 万 轮瞎猫碰死老鼠的进程，笨笨还是成功抵达终点！
笨笨和乱撞的无头苍蝇唯一的区别是：他会记录自己的每一步，成功之后可以复盘查看路径。