使用gym创建一个自定义环境

1. 环境说明

如上图，红色为起点，绿色为终点，在5*5的表格中每到达一个表格就会获得该表格对应的分数rn，agent的目标就是 找到一条路来使分数最少：）。

2. 创建环境

标准的gym环境可以分为以下几个部分，这里借鉴了gym的cliff_walking环境：

class my_env(Env):
     def __int__(self):
         pass
     def step(self.action):
         pass
     def render(self):
         pass
     def reset(self):
         pass

其中__init()用来创建动作空间，观察空间等等：

    def __init__(self):
        self.rows = 5 #行
        self.cols = 5 #列
        self.start = [0, 0] #起点
        self.goal = [3, 4] #目标点
        self.current_state = None #当前状态
        self.action_space = spaces.Discrete(3) #动作空间，离散的三个动作0，1，2
        self.observation_space = spaces.Box(low=np.array([0, 0]), high=np.array([4, 4]))#观察空间，这里为agent的坐标

step（self.action）用来定义智能体的动作：

    def step(self, action):
        new_state = deepcopy(self.current_state)
        if action == 0 : #上
            new_state[0] = max(new_state[0] - 1, 0)
        elif action == 1: #前
            new_state[1] = min(new_state[1] + 1, self.cols - 1)
        elif action == 2: #下
            new_state[0] = min(new_state[0] + 1, self.rows - 1)
        else:
            raise Exception("Invalid action")
        self.current_state = new_state


        if (self.current_state[1]-self.goal[1]) ** 2 + (self.current_state[0]- self.goal[0]) ** 2 == 0:#到达终点则游戏结束
            done = True
            reward = EachReward(current_state=self.current_state, rows=self.rows, cols=self.cols)
        else:
            done = False
            reward = EachReward(current_state=self.current_state, rows=self.rows, cols=self.cols)
            
        info = {}
        return self.current_state, reward, done, info

其中EachReward函数为自己定义的函数，具体为根据当前坐标获取分数：

import numpy as np
p = np.random.uniform(0, 1, size=(5, 5))
def EachReward(current_state,rows,cols):

    current_reward = p[current_state[0], current_state[1]]

    return current_reward

render（）为是否可视化，这里不作改动，

reset（）为重置：

    def reset(self):
        self.current_state = self.start
        return self.current_state

3. 完整代码

import gym
from gym import Env
from gym import spaces
import numpy as np
import random
from copy import deepcopy
from envs.build_problems import *

class PathPlanning(Env):
    def __init__(self):
        self.rows = 5
        self.cols = 5
        self.start = [0, 0]
        self.goal = [3, 4]
        self.current_state = None
        self.action_space = spaces.Discrete(3)
        self.observation_space = spaces.Box(low=np.array([0, 0]), high=np.array([4, 4]))
    def step(self, action):
        new_state = deepcopy(self.current_state)
        if action == 0 : #up
            new_state[0] = max(new_state[0] - 1, 0)
        elif action == 1: #forward
            new_state[1] = min(new_state[1] + 1, self.cols - 1)
        elif action == 2: #down
            new_state[0] = min(new_state[0] + 1, self.rows - 1)
        else:
            raise Exception("Invalid action")
        self.current_state = new_state


        if (self.current_state[1]-self.goal[1]) ** 2 + (self.current_state[0]- self.goal[0]) ** 2 == 0:
            done = True
            reward = 10.0
        else:
            done = False
            reward = EachReward(current_state=self.current_state, rows=self.rows, cols=self.cols)
            #print("the reward of the current state %a is %s" % (self.current_state, reward))

        info = {}
        return self.current_state, reward, done, info

    def render(self):
        pass
    def reset(self):
        self.current_state = self.start
        return self.current_state