我想為Chrome-No-Internet-Dino-Game創(chuàng)建一個(gè)AI。因此，我調(diào)整了這個(gè)Github-Repository以滿足我的需求。我使用以下公式來計(jì)算新的Q：資料來源：https://en.wikipedia.org/wiki/Q-learning我現(xiàn)在的問題是，即使在大約2.000.000次迭代之后，我的游戲分?jǐn)?shù)也沒有增加。你可以在這里找到游戲文件：https://pastebin.com/XrwQ0suJQLearning.py：import pickleimport Game_headlessimport Gameimport numpy as npfrom collections import defaultdictrewardAlive = 1rewardKill = -10000alpha = 0.2  # Learningrategamma = 0.9  # DiscountQ = defaultdict(lambda: [0, 0, 0])  # 0 = Jump / 1 = Duck / 2 = Do NothingoldState = NoneoldAction = NonegameCounter = 0gameScores = []def paramsToState(params):    cactus1X = round(params["cactus1X"] / 10) * 10    cactus2X = round(params["cactus2X"] / 10) * 10    cactus1Height = params["cactus1Height"]    cactus2Height = params["cactus2Height"]    pteraX = round(params["pteraX"] / 10) * 10    pteraY = params["pteraY"]    playerY = round(params["playerY"] / 10) * 10    gamespeed = params["gamespeed"]    return str(cactus1X) + "_" + str(cactus2X) + "_" + str(cactus1Height) + "_" + \           str(cactus2Height) + "_" + str(pteraX) + "_" + str(pteraY) + "_" + \           str(playerY) + "_" + str(gamespeed)def shouldEmulateKeyPress(params):  # 0 = Jump / 1 = Duck / 2 = Do Nothing    global oldState    global oldAction    state = paramsToState(params)    oldState = state    estReward = Q[state]    action = estReward.index(max(estReward))    if oldAction is None:        oldAction = action        return action    # Previous action was successful    # -> Update Q    prevReward = Q[oldState]    prevReward[oldAction] = (1 - alpha) * prevReward[oldAction] + \                            alpha * (rewardAlive + gamma * max(estReward))    Q[oldState] = prevReward    oldAction = action    return action在每一幀上，來自的函數(shù)調(diào)用 。然后，所述函數(shù)返回 0 表示 Jump，返回 1 表示 duck，返回 2 表示無。我嘗試調(diào)整常量，但這沒有顯示出任何效果。如果您有任何疑問，請(qǐng)隨時(shí)問我！提前感謝您！gameplay()Game_headless.pyshouldEmulateKeyPress()