class Agent(object):
def __init__(self, maze):
self.maze = Maze(maze)
self.maze.SetRewardAndNextStateDict()
self.stateNum = self.maze.m * self.maze.n
self.qTable = np.zeros([4, self.stateNum])
def GetNextState(self, state, action):
if action == 0:
return self.maze.up[state]
elif action == 1:
return self.maze.down[state]
elif action == 2:
return self.maze.left[state]
elif action == 3:
return self.maze.right[state]
import numpy as np
import random
from Maze import Maze
from Agent import Agent
mazeFileName = input("请输入迷宫文件：")
maze = np.load(mazeFileName)
mazeTraining = Maze(maze)
mazeTraining.SetRewardAndNextStateDict()
agentTraining = Agent(maze)
preTrain = input("是否使用预训练模型？（y/n）:")
if preTrain == 'y':
agentFileName = input("请输入智能体的预训练模型文件名：")
agentTraining.qTable = np.load(agentFileName)
trainingTime = int(input("请输入训练次数："))
epsilon = 1
learningRate = 0.5
gamma = 1
for i in range(trainingTime):
epsilon -= 1 / (trainingTime // 2) # epsilon递减
# 随机初始化训练起点
state = np.random.randint(agentTraining.stateNum)
while mazeTraining.GetState(state) != 1:
state = np.random.randint(agentTraining.stateNum)
time = 0
while mazeTraining.GetState(state) != 0.9:
# epsilon贪心策略
if False not in (agentTraining.qTable[:, state] == [0, 0, 0, 0]):
action = np.random.randint(4)
else:
if random.random() < epsilon:
