基於Q-Learning 的FlappyBird AI

在birdbot實現的FlappyBird基礎上訓練AI，這個FlappyBird的實現對遊戲進行了簡單的封裝，可以很方便得到遊戲的狀態來輔助演算法實現。同時可以顯示遊戲介面方便除錯，能夠看到演算法實現的效果。也可以選擇關閉遊戲介面以及聲音，這樣遊戲仍然能正常執行，一般用於訓練階段，可以減少CPU的佔用

實現參考的是SarvagyaVaish的Flappy Bird RL

Q-Learning

Q-Learning是強化學習演算法中value-based的演算法

Q即為Q（s,a）就是在某一時刻的 s 狀態下(s∈S)，採取 動作a (a∈A)動作能夠獲得收益的期望，環境會根據agent的動作反饋相應的回報reward，所以演算法的主要思想就是將State與Action構建成一張Q-table來儲存Q值，然後根據Q值來選取能夠獲得最大的收益的動作

Q-Table	a1	a2
s1	q(s1,a1)	q(s1,a2)
s2	q(s2,a1)	q(s2,a2)
s3	q(s3,a1)	q(s3,a2)

演算法流程

在更新的過程中，引入了學習速率alpha，控制先前的Q值和新的Q值之間有多少差異被保留

γ為折扣因子，0<= γ<1，γ=0表示立即回報，γ趨於1表示將來回報，γ決定時間的遠近對回報的影響程度

詳細的Q-Learning過程可以參考下面這篇

A Painless Q-learning Tutorial (一個 Q-learning 演算法的簡明教程)

FlappyBird中應用

狀態空間

• 從下方管子開始算起的垂直距離
• 從下一對管子算起的水平距離
• 鳥：死或生

動作

每一個狀態，有兩個可能的動作

• 點選一下
• 啥也不幹

獎勵

獎勵的機制完全基於鳥是否存活

• +1，如果小鳥還活著
• -1000，如果小鳥死了

流程

虛擬碼

初始化 Q = {};
while Q 未收斂：
    初始化小鳥的位置S，開始新一輪遊戲
    while S != 死亡狀態：
        使用策略π，獲得動作a=π(S) 
        使用動作a進行遊戲，獲得小鳥的新位置S',與獎勵R(S,a)
        Q[S,A] ← (1-α)*Q[S,A] + α*(R(S,a) + γ* max Q[S',a]) // 更新Q
        S ← S'

1. 觀察Flappy Bird處於什麼狀態，並執行最大化預期獎勵的行動。然後繼續執行遊戲，接著獲得下一個狀態s’

2. 觀察新的狀態s’和與之相關的獎勵：+1或者-1000

3. 根據Q Learning規則更新Q陣列

   Q[s,a] ← Q[s,a] + α (r + γ*V(s’) - Q[s,a])

4. 設定當前狀態為s’，然後重新來過

程式碼

import pyglet
import random
import pickle
import atexit
import os
from pybird.game import Game

class Bot:
    def __init__(self, game):
        self.game = game
        # constants
        self.WINDOW_HEIGHT = Game.WINDOW_HEIGHT
        self.PIPE_WIDTH = Game.PIPE_WIDTH
        # this flag is used to make sure at most one tap during
        # every call of run()
        self.tapped = False
        
        self.game.play()

        # variables for plan
        self.Q = {}
        self.alpha = 0.7
        self.explore = 100
        self.pre_s = (9999, 9999)
        self.pre_a = 'do_nothing'

        self.absolute_path = os.path.split(os.path.realpath(__file__))[0]
        self.memo = self.absolute_path + '/memo'

        if os.path.isfile(self.memo):
            _dict = pickle.load(open(self.memo))
            self.Q = _dict["Q"]
            self.game.record.iters = _dict.get("iters", 0)
            self.game.record.best_iter = _dict.get("best_iter", 0)

        def do_at_exit():
            _dict = {"Q": self.Q,
                     "iters": self.game.record.iters,
                     "best_iter": self.game.record.best_iter}
            pickle.dump(_dict, open(self.memo, 'wb'))

        atexit.register(do_at_exit)

    # this method is auto called every 0.05s by the pyglet
    def run(self):
        if self.game.state == 'PLAY':
            self.tapped = False
            # call plan() to execute your plan
            self.plan(self.get_state())
        else:
            state = self.get_state()
            bird_state = list(state['bird'])
            bird_state[2] = 'dead'
            state['bird'] = bird_state
            # do NOT allow tap
            self.tapped = True
            self.plan(state)
            # restart game
            print 'iters:',self.game.record.iters,' score:', self.game.record.get(), 'best: ', self.game.record.best_score
            self.game.record.inc_iters()
            self.game.restart()
            self.game.play()

    # get the state that robot needed
    def get_state(self):
        state = {}
        # bird's position and status(dead or alive)
        state['bird'] = (int(round(self.game.bird.x)), \
                int(round(self.game.bird.y)), 'alive')
        state['pipes'] = []
        # pipes' position
        for i in range(1, len(self.game.pipes), 2):
            p = self.game.pipes[i]
            if p.x < Game.WINDOW_WIDTH:
                # this pair of pipes shows on screen
                x = int(round(p.x))
                y = int(round(p.y))
                state['pipes'].append((x, y))
                state['pipes'].append((x, y - Game.PIPE_HEIGHT_INTERVAL))
        return state

    # simulate the click action, bird will fly higher when tapped
    # It can be called only once every time slice(every execution cycle of plan())
    def tap(self):
        if not self.tapped:
            self.game.bird.jump()
            self.tapped = True

    # That's where the robot actually works
    # NOTE Put your code here
    def plan(self, state):
        x = state['bird'][0]
        y = state['bird'][1]
        if len(state['pipes']) == 0:
            if y < self.WINDOW_HEIGHT / 2:
                self.tap()
            return
        h, v = 9999, 9999
        reward = -1000 if state['bird'][2] == 'dead' else 1
        for i in range(1, len(state['pipes']), 2):
            p = state['pipes'][i]
            if x <= p[0] + self.PIPE_WIDTH:
                h = p[0] + self.PIPE_WIDTH - x
                v = p[1] - y
                break
        scale = 10
        h /= scale
        v /= scale
        self.Q.setdefault((h, v), {'tap': 0, 'do_nothing': 0})
        self.Q.setdefault(self.pre_s, {'tap': 0, 'do_nothing': 0})
        tap_v = self.Q[(h, v)]['tap']
        nothing_v = self.Q[(h, v)]['do_nothing']
        self.Q[self.pre_s][self.pre_a] += self.alpha * (reward + max(tap_v, nothing_v) - self.Q[self.pre_s][self.pre_a])
        self.pre_s = (h, v)
        if random.randint(0, self.explore) > 100:
            self.pre_a = "do_nothing" if random.randint(0, 1) else "tap"
        else:
            tap_v = self.Q[self.pre_s]['tap']
            nothing_v = self.Q[self.pre_s]['do_nothing']
            self.pre_a = "do_nothing" if tap_v <= nothing_v else "tap"
        if self.pre_a == 'tap':
            self.tap()
        else:
            pass  

if __name__ == '__main__':
    show_window = True
    enable_sound = False
    game = Game()
    game.set_sound(enable_sound)
    bot = Bot(game)
    def update(dt):
        game.update(dt)
        bot.run()
    pyglet.clock.schedule_interval(update, Game.TIME_INTERVAL)

    if show_window:
        window = pyglet.window.Window(Game.WINDOW_WIDTH, Game.WINDOW_HEIGHT, vsync = False)
        @window.event
        def on_draw():
            window.clear()
            game.draw()
        pyglet.app.run()
    else:
        pyglet.app.run()

全部程式碼見github倉庫

參考

• Flappy Bird RL
• 如何用簡單例子講解 Q - learning 的具體過程？ - 牛阿的回答 - 知乎
• Q-Learning演算法詳解