Created
July 22, 2018 12:08
-
-
Save livoras/9e77f022a57e15d837859267270e1a79 to your computer and use it in GitHub Desktop.
Using Q-Learning to solve MountainCar-v0
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import random | |
import gym | |
import sys | |
import time | |
import pickle | |
import os | |
env = gym.make('MountainCar-v0') | |
##### | |
# { (1, 3): [actions], (3, 2): [actions], etc... } | |
qTable = {} | |
epsilon = 0.2 # 探索因子 | |
alpha = 0.5 # 学习因子 | |
gamma = 0.8 # 折扣因子 | |
MAX_EPISODE = 100000 | |
FILE_TO_SAVE = "data2" | |
lastSaveLen = 0 | |
isSuccess = 0 | |
i = 0 | |
ACTION_LEFT = 0 | |
ACTION_STAY = 1 | |
ACTION_RIGHT = 2 | |
def run(): | |
global i, isSuccess, qTable | |
qTable = loadObj(FILE_TO_SAVE) | |
while i < MAX_EPISODE: | |
state = discretizeState(env.reset()) | |
done = False | |
while not done: | |
# 渲染 | |
# if isSuccess > 500: | |
# env.render() | |
# 操作 | |
action = getActionByState(state) | |
newState, reward, done, info = env.step(action) | |
newState = discretizeState(newState) | |
updateQ(state, action, newState, reward) | |
# 切换到下一个状态了 | |
state = newState | |
# 成功了就拜拜了 | |
if newState[0] >= 0.5: | |
isSuccess = isSuccess + 1 | |
if isSuccess % 1000 == 0: | |
print("1000 Successfully! count: =>") | |
isSuccess = 0 | |
break | |
i = i + 1 | |
if isSuccess: | |
print("成功的男人!") | |
else: | |
print("失败的男人,一千回合都没有一次成功!", i) | |
def getActionByState(state): | |
hasState = state in qTable | |
# 没有状态或者要探索的时候就随机选择操作 | |
if not hasState or (random.random() <= epsilon): | |
return env.action_space.sample() | |
else: | |
# 找出所有可能的动作中最大的 Q 值的动作返回 | |
actionsQ = qTable[state] | |
maxVal = max(actionsQ) | |
return actionsQ.index(maxVal) | |
# 离散化状态,缩小状态空间 | |
def discretizeState(state): | |
return (round(state[0], 2), round(state[1], 3)) | |
# 更新Q值表 | |
def updateQ(state, action, nextState, reward): | |
global lastSaveLen | |
stateActionsQ = getActionsQByState(state) | |
nextStateActionsQ = getActionsQByState(nextState) | |
currentStateQ = stateActionsQ[action] | |
maxNextStateQ = max(nextStateActionsQ) | |
newStateQ = (1 - alpha) * currentStateQ + alpha * (reward + gamma * maxNextStateQ) | |
stateActionsQ[action] = newStateQ | |
qTable[state] = stateActionsQ | |
lenOfTable = len(qTable) | |
if (lenOfTable % 100 is 0) and (lastSaveLen != lenOfTable): | |
saveObj(qTable, FILE_TO_SAVE) | |
print("Save done, table length", lenOfTable) | |
lastSaveLen = lenOfTable | |
# time.sleep(1) | |
def getActionsQByState(state): | |
if state in qTable: | |
return qTable[state] | |
else: | |
return [0, 0, 0] | |
def saveObj(obj, name): | |
with open(name, 'wb') as f: | |
pickle.dump(obj, f, pickle.HIGHEST_PROTOCOL) | |
def loadObj(name): | |
if not os.path.exists(name): | |
return {} | |
with open(name, 'rb') as f: | |
return pickle.load(f) | |
run() |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment