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:
action = np.random.randint(4)
else:
action = np.argmax(agentTraining.qTable[:, state])
nextState = agentTraining.GetNextState(state, action)
# 更新Q表
agentTraining.qTable[action,state] =(1 - learningRate) * agentTraining.qTable[action,state] +\
learningRate * (agentTraining.maze.reward[action,state] + gamma * np.amax(agentTraining.qTable[:,nextState]))
