차이

문서의 선택한 두 판 사이의 차이를 보여줍니다.

--- elo [2018/02/27 13:49] – 만듦 rex8312
+++ elo [2024/05/08 13:34] (현재) – rex8312
@@ 줄 1: / 줄 1: @@
 ======= Elo =======
+<code python>
+import random
+import collections
+import math
+import numpy as np
+import matplotlib.pyplot as plt
+from IPython import embed
+def k_factor(player):
+    """
+    player의 k factor 계산
+    """
+    if player.elo < 1100:
+        return 25
+    elif player.elo < 2400:
+        return 15
+    else:
+        return 10
+def expected_score(player: float, opponent: float):
+    """
+    player와 opponent와 게임에서 얻을 수 있는 기대 점수(승률)
+    :param player: player의 Elo
+    :param opponent: opponent의 Elo
+    """
+    return 1 / (1 + 10 ** ((opponent - player) / 400))
+def elo(old_elo, true_score, exp_score, k):
+    """
+    player의 새로운 Elo 값 계산
+    :param old_elo: 이전 Elo
+    :param true_score: player의 실제 점수
+    :param exp_score: player의 기대 점수
+    :param k: k-factor
+    """
+    return old_elo + k * (true_score - exp_score)
+def play_game(player1, player2, n_games=1):
+    score1 = player1.play()
+    score2 = player2.play()
+    player1_score = 0.0
+    player2_score = 0.0
+    for _ in range(n_games):
+        if score1 > score2:
+            player1_score += 1.0
+            player2_score += 0.0
+        elif score1 < score2:
+            player1_score += 0.0
+            player2_score += 1.0
+        elif math.isclose(score1, score2):
+            player1_score += 0.5
+            player2_score += 0.5
+    player1_score /= n_games
+    player2_score /= n_games
+    player1_exp_score = expected_score(player1.elo, player2.elo)
+    player1_k = k_factor(player1)
+    player1.elo = elo(player1.elo, player1_score, player1_exp_score, player1_k)
+    player2_exp_score = expected_score(player2.elo, player1.elo)
+    player2_k = k_factor(player2)
+    player2.elo = elo(player2.elo, player2_score, player2_exp_score, player2_k)
+    return player1, player2
+def match_make(players):
+    return random.sample(players, 2)
+class Player:
+    def __init__(self, name, performance, init_elo=1000):
+        self.name = name
+        self._performance = performance
+        self.elo = init_elo
+    def play(self):
+        # 정규 분포보다, 균등분포가 Elo가 크게 벌어짐
+        # return random.random() + self._performance
+        # std가 커질 수록 구분이 어려움
+        return random.gauss(self._performance, 1.0)
+    def __repr__(self):
+        return f'{self.name}: {self._performance}, {int(self.elo):d}'
+def moving_average(xs, window_size=3):
+    assert len(xs) > 0
+    assert window_size > 0
+    if window_size % 2 == 0:
+        window_size += 1
+    if window_size > 1 and len(xs) > window_size:
+        weights = np.ones(window_size) / window_size
+        xs_extended = [xs]
+        for _ in range(window_size // 2):
+            xs_extended.insert(0, xs[:1])
+            xs_extended.insert(-1, xs[-1:])
+        xs_extended = np.concatenate(xs_extended)
+        return np.convolve(xs_extended, weights, mode='valid')
+    else:
+        return xs
+if __name__ == '__main__':
+    max_games = 100000
+    n_game_per_iter = 10
+    players = [Player('A', 1, init_elo=1200),
+               Player('B', 0.5), Player('C', 0.45), Player('D', 0.40), Player('E', 0.30)]
+    records = collections.deque(maxlen=max_games // n_game_per_iter)
+    for n_games in range(max_games // n_game_per_iter):
+        player1, player2 = match_make(players)
+        play_game(player1, player2, n_game_per_iter)
+        records.append([p.elo for p in players])
+    for player, elo_records in zip(players, zip(*records)):
+        elo_records = moving_average(elo_records, 100)
+        plt.plot(elo_records, label=player.name)
+    plt.legend()
+    plt.show()
+    # embed()
+</code>
+===== 참고 =====
   * https://github.com/rshk/elo/blob/master/elo.py
+  * https://github.com/djcunningham0/multielo
+  * [[trueskill|Trueskill]]