بتل کد Battle Codeدارد باز می‌شود…

Step 5: Train and Measure چالش 5 از جلسه‌ی 61، سخت

قدم ۵: تابع win_rate(games) را بنویس: games بازی با epsilon صفر انجام دهد (عامل در بازی‌های زوج g % 2 == 0 نفر اول است) و نسبت بردها را برگرداند. بعد حلقه‌ی آموزش را بنویس: ۵۰۰۰ بازی با epsilon برابر 0.3، و بعد از هر بازی learn با lr برابر 0.05 (نوبت اول هم مثل اندازه‌گیری g % 2 == 0). خروجی، بعد از خط‌های قدم‌های قبل، باید این باشد:
before training: 0.377
after training: 0.979

import random

START = 10


def moves(sticks):
    return [t for t in (1, 2, 3) if t <= sticks]


def play_game(first, second):
    sticks = START
    players = [first, second]
    turn = 0
    while True:
        t = players[turn](sticks)
        sticks = sticks - t
        if sticks <= 0:
            return turn
        turn = 1 - turn


def take_one(sticks):
    return 1


def take_most(sticks):
    return max(moves(sticks))


print(moves(10), moves(2), moves(1))
print("winner:", play_game(take_one, take_most))


def random_player(sticks):
    return random.choice(moves(sticks))


random.seed(1)
wins = [0, 0]
for g in range(1000):
    wins[play_game(random_player, random_player)] += 1
print("random vs random:", wins)


Q = {}
for s in range(1, START + 1):
    for t in moves(s):
        Q[(s, t)] = 0.0


def choose(sticks, epsilon):
    if random.random() < epsilon:
        return random.choice(moves(sticks))
    best = moves(sticks)[0]
    for t in moves(sticks):
        if Q[(sticks, t)] > Q[(sticks, best)]:
            best = t
    return best


print("Q-table size:", len(Q))
Q[(7, 3)] = 0.8
print("greedy at 7:", choose(7, 0.0), "| greedy at 5:", choose(5, 0.0))
Q[(7, 3)] = 0.0


def play_episode(epsilon, agent_first):
    sticks = START
    history = []
    agent_turn = agent_first
    while True:
        if agent_turn:
            t = choose(sticks, epsilon)
            history.append((sticks, t))
        else:
            t = random_player(sticks)
        sticks = sticks - t
        if sticks == 0:
            reward = 1 if agent_turn else -1
            return reward, history
        agent_turn = not agent_turn


def learn(history, reward, lr):
    for key in history:
        Q[key] += lr * (reward - Q[key])


random.seed(4)
reward, history = play_episode(0.2, True)
print("reward:", reward, "| agent moves:", history)
learn(history, reward, 0.5)
print("after learning:", [Q[key] for key in history])


def win_rate(games):
    wins = 0
    # play `games` greedy games (epsilon 0.0), the agent first when g is even
    return wins / games


for key in Q:
    Q[key] = 0.0
random.seed(3)
print("before training:", win_rate(1000))
# train: 5000 games with epsilon 0.3, then learn with lr 0.05
print("after training:", win_rate(1000))

راهنما و جواب این چالش · برگشت به درس جلسه‌ی 61: پروژه: عامل بازی · چالش بعدی: Step 6: The Learned Strategy

بیشتر بخوان: list comprehension در پایتون: از حلقه for تا یک خط کد