Implement correct values for rating system

Added k-factors to config.
Implemented the season 0 rating change logic.

install/default_config.yaml

@@ -19,14 +19,33 @@ variant_base_ratings:
     5p: 1700
 min_player_count: 3
 max_player_count: 5
-min_suits: 5
+
+# This adjusts the speed in rating change for players
+k-factor:
+  values:
+    # Early is applied for players with at most conditions.num_early_games games
+    early: 40
+    # This is the regular coefficient for people with a good amount of games
+    normal: 30
+    # For people with rating at least conditions.high_rating, the coefficient is adapted again,
+    high_rating: 15
+    # Controls how fast the variant ratings change
+    variants: 5
+  conditions:
+    num_early_games: 30
+    high_rating: 1700
+
+min_suits: 4
 max_suits: 6
+
 # Corresponds to game IDs from hanab.live
 starting_game_id: 1000000
 ending_game_id: 9999999
+
 # EST = Eastern Standard Time, so USA/Eastern
 starting_time: "2023-10-10 00:00:00 EST"
 ending_time: "2023-12-10 00:00:00 EST"
+
 # Any variant that contains one of these keywords will not be allowed for the league.
 excluded_variants:
   - Alternating

src/config.py

@@ -141,6 +141,36 @@ class Config:
     def excluded_variants(self) -> List[str]:
         return [var.lower() for var in self._config["excluded_variants"]]
 
+    @property
+    @check_config_attr
+    def k_factor_num_early_games(self):
+        return self._config["k-factor"]["conditions"]["num_early_games"]
+
+    @property
+    @check_config_attr
+    def k_factor_high_rating_cutoff(self):
+        return self._config["k-factor"]["conditions"]["high_rating"]
+
+    @property
+    @check_config_attr
+    def k_factor_for_few_games(self):
+        return self._config["k-factor"]["values"]["early"]
+
+    @property
+    @check_config_attr
+    def k_factor_normal(self):
+        return self._config["k-factor"]["values"]["normal"]
+
+    @property
+    @check_config_attr
+    def k_factor_for_high_rating(self):
+        return self._config["k-factor"]["values"]["high_rating"]
+
+    @property
+    @check_config_attr
+    def k_factor_for_variants(self):
+        return self._config["k-factor"]["values"]["variants"]
+
     @check_config_attr
     def variant_base_rating(self, variant_name: str, player_count: int) -> int:
         global_base_rating = self._config["variant_base_rating"]

src/constants.py

@@ -44,3 +44,6 @@ FORBIDDEN_GAME_OPTIONS = [
 # Cache time (in seconds) for history requests of players
 # In case of frequent reruns (especially during development), we do not want to stress the server too much.
 USER_HISTORY_CACHE_TIME = 5 * 60
+
+# Fraction of seeds which is assumed to be unwinnable
+UNWINNABLE_SEED_FRACTION = 0.02

src/ratings.py

@@ -5,18 +5,43 @@ from database import conn_manager
 import psycopg2.extras
 
 from log_setup import logger
+import constants
+from config import config_manager
 
 
-def rating_change(user_ratings: Dict[int, float], variant_rating: float, win: bool) -> Tuple[Dict[int, float], float]:
+def get_development_coefficient(num_games, player_rating):
+    config = config_manager.get_config()
+    if num_games <= config.k_factor_num_early_games:
+        return config.k_factor_for_few_games
+    if player_rating >= config.k_factor_high_rating_cutoff:
+        return config.k_factor_for_high_rating
+    return config.k_factor_normal
+
+
+def expected_result(player_rating, var_rating):
+    expected = (1 - constants.UNWINNABLE_SEED_FRACTION) / (1 + pow(10, (var_rating - player_rating) / 400))
+    return expected
+
+
+def compute_rating_changes(user_ratings: Dict[int, float], games_played: Dict[int, float], variant_rating: float, win: bool) -> Tuple[Dict[int, float], float]:
     """
     @param user_ratings: Mapping of user ids to ratings that played this game.
+    @param games_played: Mapping of users ids to the number of games these users played so far.
     @param variant_rating: Rating of the variant that was played
     @param win: Whether the team won the game
     @return: Mapping of user ids to their rating *changes* and *change* in variant rating
     """
-    # TODO: Implement this properly (We have not decided how this will work exactly)
-    # For now, return +1 elo for players and -1 elo for variants
-    return {user_id: 1 for user_id in user_ratings.keys()}, -1
+
+    expected_score = sum(expected_result(player_rating, variant_rating) for player_rating in user_ratings.values()) / len(user_ratings)
+    actual_score = 1 if win else 0
+    user_changes = {}
+
+    for user_id, num_games in games_played.items():
+        coefficient = get_development_coefficient(num_games, user_ratings[user_id])
+        user_changes[user_id] = coefficient * (actual_score - expected_score)
+
+    variant_change = config_manager.get_config().k_factor_for_variants * (expected_score - actual_score)
+    return user_changes, variant_change
 
 
 def next_game_to_rate():
@@ -137,29 +162,40 @@ def process_rating_of_next_game() -> bool:
         )
     league_id, num_players, score, num_suits, clue_starved, variant_id = cur.fetchone()
 
-    # Fetch game participants
-    cur.execute("SELECT game_participants.user_id FROM games "
+    # Fetch game participants and how many games they played each so far
+    cur.execute("SELECT game_participants.user_id, COUNT(games.id) "
+                "FROM game_participants "
+                "INNER JOIN games "
+                "  ON games.id = game_participants.game_id "
+                "WHERE user_id IN"
+                "  ("
+                "  SELECT game_participants.user_id FROM games "
                 "  INNER JOIN game_participants "
                 "    ON games.id = game_participants.game_id "
-                "WHERE games.id = %s",
-                (game_id,)
+                "  WHERE games.id = %s"
+                "  )"
+                "AND league_id <= %s "
+                "GROUP BY user_id",
+                (game_id, league_id)
                 )
-    user_ids = cur.fetchall()
-    if len(user_ids) != num_players:
+    games_played = {}
+    for (user_id, num_games) in cur.fetchall():
+        games_played[user_id] = num_games
+
+    if len(games_played) != num_players:
         err_msg = "Player number mismatch: Expected {} participants for game {}, but only found {} in DB: [{}]".format(
-            num_players, game_id, len(user_ids), ", ".join(user_ids)
+            num_players, game_id, len(games_played), ", ".join(games_played)
         )
         logger.error(err_msg)
         raise ValueError(err_msg)
 
     # Fetch current ratings of variant and players involved
     rating_type = utils.get_rating_type(clue_starved)
-    user_ratings = get_current_user_ratings(user_ids, rating_type)
+    user_ratings = get_current_user_ratings(list(games_played.keys()), rating_type)
     variant_rating = get_current_variant_rating(variant_id, num_players)
 
     # Calculate changes in rating
-    # TODO: If we want to use, we still have to think about how to define the K-factor and add it here
-    user_changes, variant_change = rating_change(user_ratings, variant_rating, score == 5 * num_suits)
+    user_changes, variant_change = compute_rating_changes(user_ratings, games_played, variant_rating, score == 5 * num_suits)
 
     # Update database for variants
     cur.execute("INSERT INTO variant_ratings (league_id, variant_id, num_players, change, value_after) "