Add loss minimization rating algorithm with data and results

This commit is contained in:
posij118 2024-02-05 16:35:26 +08:00
parent 4e459c4888
commit 11ec939510
4 changed files with 25985 additions and 1 deletions

25134
data/games.json Normal file

File diff suppressed because it is too large Load diff

View file

@ -0,0 +1,468 @@
{
"players": [
{
"id": 1,
"rating": 1685.2675415133688,
"username": "asaelr",
"\u03c3": 66.31240370576553
},
{
"id": 6,
"rating": 1431.613182299928,
"username": "ElenaDhynho",
"\u03c3": 29.98866321273064
},
{
"id": 7,
"rating": 1567.0774916128976,
"username": "Eliclax",
"\u03c3": 126.30005538960775
},
{
"id": 8,
"rating": 1712.039271618434,
"username": "Fafrd",
"\u03c3": 59.80163231910586
},
{
"id": 9,
"rating": 1361.5113611693616,
"username": "Feich59",
"\u03c3": 144.39407407480272
},
{
"id": 10,
"rating": 1579.6672353819206,
"username": "gw12346",
"\u03c3": 270.3079683864935
},
{
"id": 11,
"rating": 1423.2932864211616,
"username": "hakha",
"\u03c3": 85.28363174484276
},
{
"id": 12,
"rating": 1872.6154944614254,
"username": "Helana",
"\u03c3": 143.85959224548074
},
{
"id": 13,
"rating": 1527.081117882728,
"username": "hnter",
"\u03c3": 86.70665567737284
},
{
"id": 14,
"rating": 1659.6782951052335,
"username": "inseres",
"\u03c3": 54.91823796429652
},
{
"id": 15,
"rating": 1742.9546724334089,
"username": "Jay",
"\u03c3": 193.84935304947757
},
{
"id": 16,
"rating": 1667.9465689249616,
"username": "kimbifille",
"\u03c3": 15.161265503279141
},
{
"id": 17,
"rating": 1649.333225086678,
"username": "Kowalski1337",
"\u03c3": 235.3362393234084
},
{
"id": 18,
"rating": 1677.9132076067115,
"username": "macanek",
"\u03c3": 208.0527366434835
},
{
"id": 19,
"rating": 1836.1454620500044,
"username": "MarkusKahlsen",
"\u03c3": 297.71026337163414
},
{
"id": 20,
"rating": 1605.4748903814996,
"username": "Neema",
"\u03c3": 119.12840085642274
},
{
"id": 21,
"rating": 1552.9114478185877,
"username": "newduke",
"\u03c3": 109.77086600282932
},
{
"id": 22,
"rating": 1224.413480638347,
"username": "NishaNoire",
"\u03c3": 261.66467716062857
},
{
"id": 25,
"rating": 1484.5905774236187,
"username": "omegaxis",
"\u03c3": 216.00665850825928
},
{
"id": 27,
"rating": 1802.0629338345077,
"username": "pianoblook",
"\u03c3": 123.37987103512832
},
{
"id": 28,
"rating": 1664.2567844629314,
"username": "posij118",
"\u03c3": 80.37917530567991
},
{
"id": 29,
"rating": 1702.5018997700438,
"username": "purplejoe",
"\u03c3": 100.77153063473354
},
{
"id": 31,
"rating": 1892.418588622356,
"username": "rahsosprout",
"\u03c3": 104.5777908934287
},
{
"id": 32,
"rating": 1835.7346031461793,
"username": "Ramanujan",
"\u03c3": 95.89853187628516
},
{
"id": 33,
"rating": 1443.2077282559123,
"username": "ricardodd",
"\u03c3": 39.34358984367649
},
{
"id": 34,
"rating": 1040.187091444465,
"username": "RIMBarisax",
"\u03c3": 366.4360349731977
},
{
"id": 35,
"rating": 1837.7815078205804,
"username": "rz",
"\u03c3": 199.78453636449262
},
{
"id": 36,
"rating": 1696.3214849037167,
"username": "Sagnik Saha",
"\u03c3": 136.01387381633026
},
{
"id": 37,
"rating": 1605.7373915962864,
"username": "sodiumdebt",
"\u03c3": 181.85538044442802
},
{
"id": 38,
"rating": 2117.1508966520855,
"username": "spring",
"\u03c3": 57.24559998584215
},
{
"id": 39,
"rating": 1512.0428621984959,
"username": "StKildaFan",
"\u03c3": 123.53879720946766
},
{
"id": 40,
"rating": 1688.979579509527,
"username": "str8tsknacker",
"\u03c3": 139.30078280466165
},
{
"id": 41,
"rating": 1773.724951734035,
"username": "Sturm",
"\u03c3": 116.05248815088189
},
{
"id": 42,
"rating": 1510.323393769054,
"username": "TimeHoodie",
"\u03c3": 105.6566442712854
},
{
"id": 43,
"rating": 1719.6120753928153,
"username": "vEnhance",
"\u03c3": 256.98388483269576
},
{
"id": 44,
"rating": 1605.1854946464819,
"username": "vermling",
"\u03c3": 39.05915568336428
},
{
"id": 45,
"rating": 1842.1832125075996,
"username": "wateroffire",
"\u03c3": 78.80111543154953
},
{
"id": 46,
"rating": 1603.6026225502553,
"username": "WillFlame",
"\u03c3": 367.7052871728946
},
{
"id": 47,
"rating": 1747.3129095523243,
"username": "Yagami Black",
"\u03c3": 124.1498189921402
},
{
"id": 48,
"rating": 1249.7131073005326,
"username": "youisme",
"\u03c3": 119.28343113428912
},
{
"id": 51,
"rating": 1252.822074342758,
"username": "Libster",
"\u03c3": 262.48847926040565
},
{
"id": 52,
"rating": 1710.980099683423,
"username": "maxeymo",
"\u03c3": 61.64513100559617
},
{
"id": 53,
"rating": 1784.677252501476,
"username": "joano580",
"\u03c3": 290.4564102793344
},
{
"id": 54,
"rating": 1483.9387182979765,
"username": "ReaverSe",
"\u03c3": 67.78852340589508
},
{
"id": 55,
"rating": 1739.6678580930318,
"username": "benzloeb",
"\u03c3": 153.68915774257988
},
{
"id": 56,
"rating": 1554.8486203468085,
"username": "percolate",
"\u03c3": 168.9058732288808
},
{
"id": 58,
"rating": 1454.7815820698524,
"username": "Random Guy JCI",
"\u03c3": 14.92058859287638
},
{
"id": 59,
"rating": 1635.1817290999422,
"username": "aara",
"\u03c3": 113.14163640238154
},
{
"id": 60,
"rating": 1541.843324991505,
"username": "amattias",
"\u03c3": 173.9589588268638
},
{
"id": 64,
"rating": 1559.44040061339,
"username": "wtfitsnotbutter",
"\u03c3": 240.16937231126366
},
{
"id": 65,
"rating": 1441.4652215451595,
"username": "Alix_Eisenhardt",
"\u03c3": 125.35709305406003
},
{
"id": 67,
"rating": 1640.3718839938117,
"username": "Vivarus",
"\u03c3": 121.95895258848924
},
{
"id": 69,
"rating": 1582.1212409759808,
"username": "FrozenStella",
"\u03c3": 145.14226927184038
},
{
"id": 70,
"rating": 1557.096282467627,
"username": "GameConqueror",
"\u03c3": 99.69687144224555
},
{
"id": 71,
"rating": 1662.4377474617993,
"username": "seungapark",
"\u03c3": 145.15724952841202
},
{
"id": 72,
"rating": 1508.3443601141673,
"username": "sinuni_hung",
"\u03c3": 61.81213974983676
},
{
"id": 73,
"rating": 1803.5327953459553,
"username": "TheDaniMan",
"\u03c3": 134.82661666333226
},
{
"id": 79,
"rating": 1482.0644773366396,
"username": "Kaznad",
"\u03c3": 405.28035311445757
},
{
"id": 80,
"rating": 1086.1355090178276,
"username": "Kernel",
"\u03c3": 211.39242310002254
},
{
"id": 81,
"rating": 1290.2977493468493,
"username": "Le Codex",
"\u03c3": 128.6528429208025
},
{
"id": 82,
"rating": 1554.5865377995306,
"username": "Nipalup",
"\u03c3": 186.7194364889647
}
],
"variants": [
{
"id": 0,
"rating": 1344.571218042761,
"num_players": 3,
"name": "No Variant",
"num_suits": 5,
"\u03c3": 31.335534554470627
},
{
"id": 1,
"rating": 1380.3108266495492,
"num_players": 3,
"name": "6 Suits",
"num_suits": 6,
"\u03c3": 33.55653633693724
},
{
"id": 51,
"rating": 1751.8509769680584,
"num_players": 3,
"name": "Clue Starved (6 Suits)",
"num_suits": 6,
"\u03c3": 11.424872579830076
},
{
"id": 52,
"rating": 1677.1056853935763,
"num_players": 3,
"name": "Clue Starved (5 Suits)",
"num_suits": 5,
"\u03c3": 40.42763640571963
},
{
"id": 0,
"rating": 1436.6029781659563,
"num_players": 4,
"name": "No Variant",
"num_suits": 5,
"\u03c3": 45.77891871651097
},
{
"id": 1,
"rating": 1408.254084838167,
"num_players": 4,
"name": "6 Suits",
"num_suits": 6,
"\u03c3": 14.953524984939428
},
{
"id": 51,
"rating": 1820.7964465875705,
"num_players": 4,
"name": "Clue Starved (6 Suits)",
"num_suits": 6,
"\u03c3": 81.2114821701358
},
{
"id": 52,
"rating": 1719.526642788024,
"num_players": 4,
"name": "Clue Starved (5 Suits)",
"num_suits": 5,
"\u03c3": 38.5663698641606
},
{
"id": 0,
"rating": 1373.7429515570884,
"num_players": 5,
"name": "No Variant",
"num_suits": 5,
"\u03c3": 72.00085092075848
},
{
"id": 1,
"rating": 1522.642598263524,
"num_players": 5,
"name": "6 Suits",
"num_suits": 6,
"\u03c3": 36.55838248550461
},
{
"id": 51,
"rating": 1895.801292104732,
"num_players": 5,
"name": null,
"num_suits": 6,
"\u03c3": 14.915817466505452
},
{
"id": 52,
"rating": 1968.4174574449692,
"num_players": 5,
"name": "Clue Starved (5 Suits)",
"num_suits": 5,
"\u03c3": 50.09621953978423
}
]
}

View file

@ -28,7 +28,7 @@ DB_TABLE_NAMES = [
DATABASE_SCHEMA_PATH = 'install/database_schema.sql'
DEFAULT_DB_CONFIG_PATH = 'install/default_db_config.yaml'
DEFAULT_CONFIG_PATH = 'install/default_config.yaml'
DEFAULT_CONFIG_PATH = '../install/default_config.yaml'
VARIANTS_JSON_URL = 'https://raw.githubusercontent.com/Hanabi-Live/hanabi-live/main/packages/game/src/json/variants.json'

382
src/minimize_loss.py Normal file
View file

@ -0,0 +1,382 @@
from typing import Iterable, List
from config import config_manager
from constants import UNWINNABLE_SEED_FRACTION
from dataclasses import dataclass
import json
import datetime
import numpy as np
from numpy.typing import NDArray
from scipy.optimize import minimize, Bounds
MAX_RATING_DIFF = 1200
RATING_PRIOR = 1600
# This allows us to do integer optimization on more fine-grained grid.
LCM_FACTOR = 60
NUM_RANDOM_INITS = 50
NUM_CV_ITERS = 50
VARIANT_DEFAULT_RATINGS = np.array([1400, 1400, 1800, 1800]).reshape((1, 4))
VARIANT_NUM_PLAYERS_MODIFIERS = np.array([0, 0, 100]).reshape((3, 1))
# As of season 1, this is [1400 1400 1800 1800 1400 1400 1800 1800 1500 1500 1900 1900].
VARIANT_RATING_PRIORS = np.ravel(
VARIANT_DEFAULT_RATINGS + VARIANT_NUM_PLAYERS_MODIFIERS
)
@dataclass
class GlobalInfo:
rated_ids: Iterable[int]
variant_ids: Iterable[int]
player_counts: Iterable[int]
user_names: Iterable[str]
game_counts: Iterable[int]
@property
def rating_list_length(self):
return len(self.rated_ids) + len(self.variant_ids) * len(self.player_counts)
@property
def rated_id_indices_in_rating_list(self):
return {id: index for index, id in dict(enumerate(self.rated_ids)).items()}
@dataclass
class GameRow:
game_id: int
num_players: int
users: List[str]
user_ids: List[int]
user_rating_changes: List[float]
user_ratings_after: List[float]
variant_id: int
variant_name: str
num_suits: id
rating_type: int
seed: str
score: int
num_turns: int
datetime_finished: datetime.datetime
league_id: int
num_bdrs: int
num_crits_lost: int
game_outcomes: List[str]
variant_rating_change: float
variant_rating_after: float
def calculate_loss(
rating_list: Iterable[int],
game_list: Iterable[GameRow],
global_info: GlobalInfo,
p_win_lookup_table: NDArray[np.float32],
player_rating_priors: Iterable[int],
player_prior_weight: float,
variant_rating_priors: Iterable[int],
variant_prior_weight: float,
):
loss = np.float32(0)
for game in game_list:
team_ratings = [
rating_list[global_info.rated_id_indices_in_rating_list[user_id]]
for user_id in game["user_ids"]
]
team_rating = sum(team_ratings) / len(team_ratings)
variant_rating = rating_list[
calculate_variant_index_in_rating_list(
len(game["users"]), game["variant_id"], global_info
)
]
rating_diff = int(np.round(LCM_FACTOR * (team_rating - variant_rating)))
p_win = calculate_p_win(rating_diff, p_win_lookup_table)
game_won = game["game_outcomes"].count("Win") > 0
loss += calculate_loss_for_one_game(p_win, game_won)
# Add two dummy single-player games for each player with prior_weight * square-root of number of games weight - one won and lost
for player_rating_prior, rated_id, game_count in zip(
player_rating_priors, global_info.rated_ids, global_info.game_counts
):
team_rating = rating_list[global_info.rated_id_indices_in_rating_list[rated_id]]
variant_rating = player_rating_prior
rating_diff = int(np.round(LCM_FACTOR * (team_rating - variant_rating)))
p_win = calculate_p_win(rating_diff, p_win_lookup_table)
loss += (
np.sqrt(game_count)
* player_prior_weight
* calculate_loss_for_one_game(p_win, True)
)
loss += (
np.sqrt(game_count)
* player_prior_weight
* calculate_loss_for_one_game(p_win, False)
)
# Add two dummy single-player games for each player with prior_weight * square-root of number of games weight - one won and lost
for variant_rating, variant_rating_prior in zip(
rating_list[len(global_info.rated_ids) :],
variant_rating_priors,
):
team_rating = variant_rating_prior
rating_diff = int(np.round(LCM_FACTOR * (team_rating - variant_rating)))
p_win = calculate_p_win(rating_diff, p_win_lookup_table)
loss += (
np.sqrt(game_count)
* variant_prior_weight
* calculate_loss_for_one_game(p_win, True)
)
loss += (
np.sqrt(game_count)
* variant_prior_weight
* calculate_loss_for_one_game(p_win, False)
)
return loss
def calculate_loss_gradient(rating_list, *args):
y = calculate_loss(rating_list, *args)
eye = np.eye(len(rating_list), dtype=int)
gradient = [
calculate_loss(rating_list + eye[i], *args) - y for i in range(len(rating_list))
]
return gradient
def calculate_minloss_ratings(
game_list: List[GameRow],
global_info: GlobalInfo,
p_win_lookup_table: NDArray[np.float32],
player_rating_priors: int | Iterable[int] = RATING_PRIOR,
player_random_init_stdev: int = 0,
player_prior_weight=0.5,
variant_rating_priors: int | Iterable[int] = RATING_PRIOR,
variant_prior_weight=0.5,
variant_random_init_stdev: int = 0,
):
player_rating_priors = np.broadcast_to(player_rating_priors, len(rated_ids))
variant_rating_priors = np.broadcast_to(
variant_rating_priors, len(variant_ids) * len(player_counts)
)
prior_rating_list = np.concatenate(
(
player_rating_priors
+ np.random.normal(0, player_random_init_stdev, len(player_rating_priors)),
variant_rating_priors
+ np.random.normal(
0, variant_random_init_stdev, len(variant_rating_priors)
),
),
dtype=np.float32,
)
rating_list = minimize(
calculate_loss,
prior_rating_list,
(
game_list,
global_info,
p_win_lookup_table,
player_rating_priors,
player_prior_weight,
variant_rating_priors,
variant_prior_weight,
),
bounds=Bounds(RATING_PRIOR - MAX_RATING_DIFF, RATING_PRIOR + MAX_RATING_DIFF),
jac=calculate_loss_gradient,
tol=0.0001,
)
return rating_list
def calculate_p_win_lookup_table(rating_diff_indices: Iterable[int]):
lookup_table = np.full(2 * MAX_RATING_DIFF * LCM_FACTOR + 1, 0.5, dtype=np.float32)
for rating_diff_index in rating_diff_indices:
p_win = calculate_p_win(rating_diff_index, None)
lookup_table[rating_diff_index] = p_win
return lookup_table
def calculate_p_win(
rating_diff_index: int, lookup_table: NDArray[np.float32] | None = None
):
if lookup_table is None:
return (1 - UNWINNABLE_SEED_FRACTION) * (
1 / (1 + 10 ** (-(rating_diff_index / LCM_FACTOR) / 400))
)
else:
return lookup_table[rating_diff_index]
# Cross-entropy loss
def calculate_loss_for_one_game(p_win: np.float32, game_won: bool):
if game_won:
loss = -np.log(p_win)
else:
loss = -np.log(1 - p_win)
return loss
def calculate_variant_index_in_rating_list(
player_count: int, variant_id: int, global_info: GlobalInfo
):
return (
len(global_info.rated_ids)
+ global_info.player_counts.index(player_count) * len(global_info.variant_ids)
+ global_info.variant_ids.index(variant_id)
)
def calculate_player_count_and_variant_id_from_variant_index(
index: int, global_info: GlobalInfo
):
num_players = global_info.player_counts[index // len(global_info.variant_ids)]
variant_id = global_info.variant_ids[index % len(global_info.variant_ids)]
return num_players, variant_id
def write_data(random_init_rating_lists, cv_rating_lists):
with open("../results/min_loss_rating_list.json", "w") as f:
rating_list = np.average(random_init_rating_lists, axis=0)
rating_stdevs = np.std(cv_rating_lists, axis=0, ddof=1)
players = []
for rated_id, user_name, rating, rating_stdev in zip(
global_info.rated_ids, global_info.user_names, rating_list, rating_stdevs
):
players.append(
{
"id": rated_id,
"rating": rating,
"username": user_name,
"stdev": rating_stdev,
}
)
variants = []
for i, (rating, rating_stdev) in enumerate(
zip(
rating_list[len(global_info.rated_ids) :],
rating_stdevs[len(global_info.rated_ids) :],
)
):
(
num_players,
variant_id,
) = calculate_player_count_and_variant_id_from_variant_index(i, global_info)
variant_name = None
for game in game_list:
if (
len(game["users"]) == num_players
and game["variant_id"] == variant_id
):
variant_name = game["variant_name"]
num_suits = game["num_suits"]
variants.append(
{
"id": variant_id,
"rating": rating,
"num_players": num_players,
"name": variant_name,
"num_suits": num_suits,
"stdev": rating_stdev,
}
)
s = json.dumps({"players": players, "variants": variants})
f.write(s)
if __name__ == "__main__":
with open("../data/games.json") as game_data:
game_list = json.loads(game_data.read())
p_win_lookup_table = calculate_p_win_lookup_table(
np.arange(-MAX_RATING_DIFF * LCM_FACTOR, MAX_RATING_DIFF * LCM_FACTOR + 1)
)
config = config_manager.get_config()
rated_ids_dupes = [ID for game in game_list for ID in game["user_ids"]]
rated_ids = list(set(rated_ids_dupes))
variant_ids_dupes = [game["variant_id"] for game in game_list]
variant_ids = list(set(variant_ids_dupes))
nums_players = [game["num_players"] for game in game_list]
player_counts = list(range(config.min_player_count, config.max_player_count + 1))
game_counts = np.concatenate(
[rated_ids_dupes.count(ID) for ID in rated_ids],
np.zeros(len(variant_ids) * len(player_counts)),
)
user_names = list(np.full((len(rated_ids),), ""))
for game in game_list:
for user_id, user_name in zip(game["user_ids"], game["users"]):
user_names[rated_ids.index(user_id)] = user_name
global_info = GlobalInfo(
rated_ids, variant_ids, player_counts, user_names, game_counts
)
for variant_id, num_players in zip(variant_ids_dupes, nums_players):
game_counts[
len(rated_ids)
+ calculate_variant_index_in_rating_list(
num_players, variant_id, global_info
)
] += 1
global_info.game_counts = game_counts
random_init_rating_lists = []
cv_rating_lists = []
for _ in range(NUM_RANDOM_INITS):
random_init_rating_lists.append(
calculate_minloss_ratings(
game_list,
global_info,
p_win_lookup_table,
variant_rating_priors=VARIANT_RATING_PRIORS,
player_random_init_stdev=200,
variant_random_init_stdev=200,
).x
)
for _ in range(NUM_CV_ITERS):
cv_game_list = np.random.choice(game_list, len(game_list), replace=True)
cv_rating_lists.append(
calculate_minloss_ratings(
cv_game_list,
global_info,
p_win_lookup_table,
variant_rating_priors=VARIANT_RATING_PRIORS,
player_random_init_stdev=200,
variant_random_init_stdev=200,
).x
)
write_data(random_init_rating_lists, cv_rating_lists)