Fix AI swapping high cards into unknown positions
Prevent CPU players from swapping 8+ value cards (8, 9, 10, J, Q) into face-down positions, which is statistically bad since expected hidden card value is ~4.5. Fixes applied: - Add value threshold (7) to unpredictability random swap path - Restrict comeback bonus to cards with value < 8 - Reduce speculative wolfpack Jack bonus from 6x to 2x aggression - Add safety filter to remove hidden positions for 8+ cards - Fix endgame logic to discard 8+ instead of forcing swap into hidden - Skip hidden positions in denial candidate list for 8+ cards - Add swapped_high_into_unknown tracking to SimulationStats Reduces "swapped 8+ into unknown" dumb moves from ~85 per 200 games to ~6 per 500 games (0.054% rate, down from ~2%). Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
parent
9b53e51aa3
commit
4ad508f84f
61
server/ai.py
61
server/ai.py
@ -1018,10 +1018,13 @@ class GolfAI:
|
||||
score += wolfpack_bonus
|
||||
ai_log(f" Wolfpack pursuit: completing 2nd Jack pair! +{wolfpack_bonus:.1f}")
|
||||
elif not partner_card.face_up:
|
||||
# Partner unknown, Jack could pair
|
||||
wolfpack_bonus = 6 * profile.aggression
|
||||
# Partner unknown - speculative wolfpack pursuit
|
||||
# Probability of unknown card being Jack is very low (~3%)
|
||||
# Expected value of swapping Jack into unknown is negative
|
||||
# Only give small bonus - not enough to override negative point_gain
|
||||
wolfpack_bonus = 2 * profile.aggression
|
||||
score += wolfpack_bonus
|
||||
ai_log(f" Wolfpack pursuit: Jack with unknown partner +{wolfpack_bonus:.1f}")
|
||||
ai_log(f" Wolfpack pursuit (speculative): +{wolfpack_bonus:.1f}")
|
||||
elif visible_jacks >= 1 and partner_card.face_up and partner_card.rank == Rank.JACK:
|
||||
# Completing first Jack pair while having other Jacks
|
||||
wolfpack_bonus = 8 * profile.aggression
|
||||
@ -1029,8 +1032,10 @@ class GolfAI:
|
||||
ai_log(f" Wolfpack pursuit: first Jack pair +{wolfpack_bonus:.1f}")
|
||||
|
||||
# 4d. COMEBACK AGGRESSION - Boost reveal bonus when behind in late game
|
||||
# Only for cards that aren't objectively bad (value < 8)
|
||||
# Don't incentivize locking in 8, 9, 10, J, Q just to "go out faster"
|
||||
standings_pressure = get_standings_pressure(player, game)
|
||||
if standings_pressure > 0.3 and not current_card.face_up:
|
||||
if standings_pressure > 0.3 and not current_card.face_up and drawn_value < 8:
|
||||
# Behind in standings - boost incentive to reveal and play faster
|
||||
comeback_bonus = standings_pressure * 3 * profile.aggression
|
||||
score += comeback_bonus
|
||||
@ -1222,9 +1227,17 @@ class GolfAI:
|
||||
ai_log(f" >> UNPREDICTABLE: randomly chose position {last_pos} (projected {projected})")
|
||||
return last_pos
|
||||
else:
|
||||
choice = random.choice(face_down)
|
||||
ai_log(f" >> UNPREDICTABLE: randomly chose position {choice}")
|
||||
return choice
|
||||
# Only allow random swaps for cards that aren't objectively bad
|
||||
# Cards 8+ are too bad to randomly put into unknowns
|
||||
# (Expected value of hidden card is ~4.5)
|
||||
UNPREDICTABLE_MAX_VALUE = 7
|
||||
if drawn_value <= UNPREDICTABLE_MAX_VALUE:
|
||||
choice = random.choice(face_down)
|
||||
ai_log(f" >> UNPREDICTABLE: randomly chose position {choice} (value {drawn_value} <= {UNPREDICTABLE_MAX_VALUE})")
|
||||
return choice
|
||||
else:
|
||||
ai_log(f" >> UNPREDICTABLE: blocked - value {drawn_value} > {UNPREDICTABLE_MAX_VALUE} threshold")
|
||||
# Fall through to normal scoring logic
|
||||
|
||||
# Calculate score for each position
|
||||
position_scores: list[tuple[int, float]] = []
|
||||
@ -1249,6 +1262,24 @@ class GolfAI:
|
||||
# Filter to positive scores only
|
||||
positive_scores = [(p, s) for p, s in position_scores if s > 0]
|
||||
|
||||
# SAFETY: Never swap high cards (8+) into hidden positions
|
||||
# This is objectively bad since expected hidden value is ~4.5
|
||||
# Exception: creating a visible pair (partner face-up and matches)
|
||||
if drawn_value >= 8:
|
||||
safe_positive = []
|
||||
for pos, score in positive_scores:
|
||||
card = player.cards[pos]
|
||||
partner_pos = get_column_partner_position(pos)
|
||||
partner = player.cards[partner_pos]
|
||||
creates_pair = partner.face_up and partner.rank == drawn_card.rank
|
||||
|
||||
if card.face_up or creates_pair:
|
||||
safe_positive.append((pos, score))
|
||||
else:
|
||||
ai_log(f" SAFETY: rejecting pos {pos} - high card ({drawn_value}) into hidden")
|
||||
|
||||
positive_scores = safe_positive
|
||||
|
||||
best_pos: Optional[int] = None
|
||||
best_score = 0.0
|
||||
|
||||
@ -1368,11 +1399,16 @@ class GolfAI:
|
||||
ai_log(f" >> FINAL SAFETY: swapping into visible pos {worst_visible_pos} "
|
||||
f"(drawn {drawn_value} < worst visible {worst_visible_val})")
|
||||
best_pos = worst_visible_pos
|
||||
elif drawn_value >= 8:
|
||||
# Drawn card is terrible (8+) - better to discard and flip the unknown
|
||||
# Don't lock in a guaranteed bad card
|
||||
ai_log(f" >> FINAL SAFETY: discarding bad card ({drawn_value}), will flip unknown")
|
||||
best_pos = None # Discard
|
||||
else:
|
||||
# Drawn card is terrible and worse than all visible cards
|
||||
# Swap into hidden position anyway - known bad is better than unknown
|
||||
# Drawn card is mediocre but not terrible - swap into hidden
|
||||
# known mediocre is better than unknown
|
||||
ai_log(f" >> FINAL SAFETY: forcing swap into hidden pos {last_pos} "
|
||||
f"(all options bad, but known > unknown)")
|
||||
f"(drawn value {drawn_value} is acceptable)")
|
||||
best_pos = last_pos
|
||||
|
||||
# OPPONENT DENIAL CHECK: Before discarding, check if this would help next player
|
||||
@ -1396,6 +1432,9 @@ class GolfAI:
|
||||
card = player.cards[pos]
|
||||
if not card.face_up:
|
||||
# Swapping into face-down: cost is drawn_value (we keep it)
|
||||
# Skip hidden positions for high cards (8+) - too costly
|
||||
if drawn_value >= 8:
|
||||
continue # Never swap 8+ into hidden for denial
|
||||
cost = drawn_value
|
||||
denial_candidates.append((pos, cost, "hidden"))
|
||||
else:
|
||||
@ -1699,7 +1738,7 @@ async def process_cpu_turn(
|
||||
) -> None:
|
||||
"""Process a complete turn for a CPU player."""
|
||||
import asyncio
|
||||
from game_log import get_logger
|
||||
from services.game_logger import get_logger
|
||||
|
||||
profile = get_profile(cpu_player.id)
|
||||
if not profile:
|
||||
|
||||
@ -28,7 +28,9 @@ from ai import (
|
||||
filter_bad_pair_positions, get_column_partner_position
|
||||
)
|
||||
from game import Rank
|
||||
from game_log import GameLogger
|
||||
|
||||
# Note: Simulations run standalone without PostgreSQL database logging.
|
||||
# In-memory SimulationStats provides all the analysis needed for bulk runs.
|
||||
|
||||
|
||||
# Named rule presets for quick configuration
|
||||
@ -121,6 +123,7 @@ class SimulationStats:
|
||||
self.took_bad_card_without_pair = 0
|
||||
self.paired_negative_cards = 0
|
||||
self.swapped_good_for_bad = 0
|
||||
self.swapped_high_into_unknown = 0 # Cards 8+ swapped into face-down position
|
||||
self.total_opportunities = 0 # Total decision points
|
||||
|
||||
def record_game(self, game: Game, winner_name: str):
|
||||
@ -158,6 +161,8 @@ class SimulationStats:
|
||||
self.paired_negative_cards += 1
|
||||
elif move_type == "swapped_good_for_bad":
|
||||
self.swapped_good_for_bad += 1
|
||||
elif move_type == "swapped_high_into_unknown":
|
||||
self.swapped_high_into_unknown += 1
|
||||
|
||||
def record_opportunity(self):
|
||||
"""Record a decision opportunity for rate calculation."""
|
||||
@ -172,7 +177,8 @@ class SimulationStats:
|
||||
self.discarded_kings +
|
||||
self.took_bad_card_without_pair +
|
||||
self.paired_negative_cards +
|
||||
self.swapped_good_for_bad
|
||||
self.swapped_good_for_bad +
|
||||
self.swapped_high_into_unknown
|
||||
)
|
||||
if self.total_opportunities == 0:
|
||||
return 0.0
|
||||
@ -230,6 +236,7 @@ class SimulationStats:
|
||||
lines.append(" Mistakes (should be < 0.1%):")
|
||||
lines.append(f" Discarded Kings: {self.discarded_kings}")
|
||||
lines.append(f" Swapped good for bad: {self.swapped_good_for_bad}")
|
||||
lines.append(f" Swapped 8+ into unknown: {self.swapped_high_into_unknown}")
|
||||
|
||||
return "\n".join(lines)
|
||||
|
||||
@ -251,8 +258,6 @@ def run_cpu_turn(
|
||||
game: Game,
|
||||
player: Player,
|
||||
profile: CPUProfile,
|
||||
logger: Optional[GameLogger],
|
||||
game_id: Optional[str],
|
||||
stats: SimulationStats
|
||||
) -> str:
|
||||
"""Run a single CPU turn synchronously. Returns action taken."""
|
||||
@ -291,39 +296,45 @@ def run_cpu_turn(
|
||||
if not has_pair_potential and not has_worse_to_replace:
|
||||
stats.record_dumb_move("took_bad_without_pair")
|
||||
|
||||
# Log draw decision
|
||||
if logger and game_id:
|
||||
reason = f"took {discard_top.rank.value} from discard" if take_discard else "drew from deck"
|
||||
logger.log_move(
|
||||
game_id=game_id,
|
||||
player=player,
|
||||
is_cpu=True,
|
||||
action=action,
|
||||
card=drawn,
|
||||
game=game,
|
||||
decision_reason=reason,
|
||||
)
|
||||
|
||||
# Decide whether to swap or discard
|
||||
swap_pos = GolfAI.choose_swap_or_discard(drawn, player, profile, game)
|
||||
ai_chose_swap = swap_pos is not None # Track if AI made this decision vs fallback
|
||||
|
||||
# If drawn from discard, must swap
|
||||
if swap_pos is None and game.drawn_from_discard:
|
||||
face_down = [i for i, c in enumerate(player.cards) if not c.face_up]
|
||||
if face_down:
|
||||
# Use filter to avoid bad pairs with negative cards
|
||||
safe_positions = filter_bad_pair_positions(face_down, drawn, player, game.options)
|
||||
swap_pos = random.choice(safe_positions)
|
||||
else:
|
||||
# Find worst card using house rules
|
||||
worst_pos = 0
|
||||
worst_val = -999
|
||||
for i, c in enumerate(player.cards):
|
||||
drawn_val = get_ai_card_value(drawn, game.options)
|
||||
|
||||
# First, check if there's a visible card WORSE than what we drew
|
||||
# (prefer swapping visible bad cards over face-down unknowns)
|
||||
worst_visible_pos = None
|
||||
worst_visible_val = drawn_val # Only consider cards worse than drawn
|
||||
for i, c in enumerate(player.cards):
|
||||
if c.face_up:
|
||||
card_val = get_ai_card_value(c, game.options)
|
||||
if card_val > worst_val:
|
||||
worst_val = card_val
|
||||
worst_pos = i
|
||||
swap_pos = worst_pos
|
||||
if card_val > worst_visible_val:
|
||||
worst_visible_val = card_val
|
||||
worst_visible_pos = i
|
||||
|
||||
if worst_visible_pos is not None:
|
||||
# Found a visible card worse than drawn - swap with it
|
||||
swap_pos = worst_visible_pos
|
||||
else:
|
||||
# No visible card worse than drawn - must use face-down
|
||||
face_down = [i for i, c in enumerate(player.cards) if not c.face_up]
|
||||
if face_down:
|
||||
# Use filter to avoid bad pairs with negative cards
|
||||
safe_positions = filter_bad_pair_positions(face_down, drawn, player, game.options)
|
||||
swap_pos = random.choice(safe_positions)
|
||||
else:
|
||||
# All cards face-up, find worst card overall
|
||||
worst_pos = 0
|
||||
worst_val = -999
|
||||
for i, c in enumerate(player.cards):
|
||||
card_val = get_ai_card_value(c, game.options)
|
||||
if card_val > worst_val:
|
||||
worst_val = card_val
|
||||
worst_pos = i
|
||||
swap_pos = worst_pos
|
||||
|
||||
# Record this as a decision opportunity for dumb move rate calculation
|
||||
stats.record_opportunity()
|
||||
@ -362,6 +373,24 @@ def run_cpu_turn(
|
||||
if not creates_pair and not is_denial_move:
|
||||
stats.record_dumb_move("swapped_good_for_bad")
|
||||
|
||||
# Check for dumb move: swapping high card into unknown position
|
||||
# Cards 8+ (8, 9, 10, J, Q) should never be swapped into face-down positions
|
||||
# since expected value of hidden card is only ~4.5
|
||||
# Exception: pairing, denial moves, or forced swaps from discard
|
||||
if not old_card.face_up and drawn_val >= 8:
|
||||
if not creates_pair and not is_denial_move:
|
||||
# Only count as dumb if:
|
||||
# 1. AI actively chose this (not fallback from forced discard swap)
|
||||
# 2. OR if drawn from discard but a worse visible card existed
|
||||
worse_visible_exists = has_worse_visible_card(player, drawn_val, game.options)
|
||||
|
||||
if ai_chose_swap:
|
||||
# AI chose to swap 8+ into hidden - this is dumb
|
||||
stats.record_dumb_move("swapped_high_into_unknown")
|
||||
elif game.drawn_from_discard and worse_visible_exists:
|
||||
# Fallback chose hidden when worse visible existed - also dumb
|
||||
stats.record_dumb_move("swapped_high_into_unknown")
|
||||
|
||||
# Check for dumb move: creating bad pair with negative card
|
||||
if (partner.face_up and
|
||||
partner.rank == drawn.rank and
|
||||
@ -375,18 +404,6 @@ def run_cpu_turn(
|
||||
game.swap_card(player.id, swap_pos)
|
||||
action = "swap"
|
||||
stats.record_turn(player.name, action)
|
||||
|
||||
if logger and game_id:
|
||||
logger.log_move(
|
||||
game_id=game_id,
|
||||
player=player,
|
||||
is_cpu=True,
|
||||
action="swap",
|
||||
card=drawn,
|
||||
position=swap_pos,
|
||||
game=game,
|
||||
decision_reason=f"swapped {drawn.rank.value} for {old_card.rank.value} at pos {swap_pos}",
|
||||
)
|
||||
else:
|
||||
# Check for dumb moves: discarding excellent cards
|
||||
if drawn.rank == Rank.JOKER:
|
||||
@ -400,41 +417,16 @@ def run_cpu_turn(
|
||||
action = "discard"
|
||||
stats.record_turn(player.name, action)
|
||||
|
||||
if logger and game_id:
|
||||
logger.log_move(
|
||||
game_id=game_id,
|
||||
player=player,
|
||||
is_cpu=True,
|
||||
action="discard",
|
||||
card=drawn,
|
||||
game=game,
|
||||
decision_reason=f"discarded {drawn.rank.value}",
|
||||
)
|
||||
|
||||
if game.flip_on_discard:
|
||||
flip_pos = GolfAI.choose_flip_after_discard(player, profile)
|
||||
game.flip_and_end_turn(player.id, flip_pos)
|
||||
|
||||
if logger and game_id:
|
||||
flipped = player.cards[flip_pos]
|
||||
logger.log_move(
|
||||
game_id=game_id,
|
||||
player=player,
|
||||
is_cpu=True,
|
||||
action="flip",
|
||||
card=flipped,
|
||||
position=flip_pos,
|
||||
game=game,
|
||||
decision_reason=f"flipped position {flip_pos}",
|
||||
)
|
||||
|
||||
return action
|
||||
|
||||
|
||||
def run_game(
|
||||
players_with_profiles: list[tuple[Player, CPUProfile]],
|
||||
options: GameOptions,
|
||||
logger: Optional[GameLogger],
|
||||
stats: SimulationStats,
|
||||
verbose: bool = False
|
||||
) -> tuple[str, int]:
|
||||
@ -455,15 +447,6 @@ def run_game(
|
||||
|
||||
game.start_game(num_decks=1, num_rounds=1, options=options)
|
||||
|
||||
# Log game start
|
||||
game_id = None
|
||||
if logger:
|
||||
game_id = logger.log_game_start(
|
||||
room_code="SIM",
|
||||
num_players=len(players_with_profiles),
|
||||
options=options
|
||||
)
|
||||
|
||||
# Do initial flips for all players
|
||||
if options.initial_flips > 0:
|
||||
for player, profile in players_with_profiles:
|
||||
@ -480,17 +463,13 @@ def run_game(
|
||||
break
|
||||
|
||||
profile = profiles[current.id]
|
||||
action = run_cpu_turn(game, current, profile, logger, game_id, stats)
|
||||
action = run_cpu_turn(game, current, profile, stats)
|
||||
|
||||
if verbose and turn_count % 10 == 0:
|
||||
print(f" Turn {turn_count}: {current.name} - {action}")
|
||||
|
||||
turn_count += 1
|
||||
|
||||
# Log game end
|
||||
if logger and game_id:
|
||||
logger.log_game_end(game_id)
|
||||
|
||||
# Find winner
|
||||
winner = min(game.players, key=lambda p: p.total_score)
|
||||
stats.record_game(game, winner.name)
|
||||
@ -523,7 +502,6 @@ def run_simulation(
|
||||
print(f"Rules: {rules_desc}")
|
||||
print("=" * 50)
|
||||
|
||||
logger = GameLogger()
|
||||
stats = SimulationStats()
|
||||
|
||||
for i in range(num_games):
|
||||
@ -533,7 +511,7 @@ def run_simulation(
|
||||
names = [p.name for p, _ in players]
|
||||
print(f"\nGame {i+1}/{num_games}: {', '.join(names)}")
|
||||
|
||||
winner, score = run_game(players, options, logger, stats, verbose=False)
|
||||
winner, score = run_game(players, options, stats, verbose=False)
|
||||
|
||||
if verbose:
|
||||
print(f" Winner: {winner} (score: {score})")
|
||||
@ -541,13 +519,6 @@ def run_simulation(
|
||||
print("\n")
|
||||
print(stats.report())
|
||||
|
||||
print("\n" + "=" * 50)
|
||||
print("ANALYSIS")
|
||||
print("=" * 50)
|
||||
print("\nRun analysis with:")
|
||||
print(" python game_analyzer.py blunders")
|
||||
print(" python game_analyzer.py summary")
|
||||
|
||||
return stats
|
||||
|
||||
|
||||
@ -571,7 +542,6 @@ def run_detailed_game(num_players: int = 4, options: Optional[GameOptions] = Non
|
||||
print(f"Rules: {rules_desc}")
|
||||
print("=" * 50)
|
||||
|
||||
logger = GameLogger()
|
||||
stats = SimulationStats()
|
||||
|
||||
players_with_profiles = create_cpu_players(num_players)
|
||||
@ -586,12 +556,6 @@ def run_detailed_game(num_players: int = 4, options: Optional[GameOptions] = Non
|
||||
|
||||
game.start_game(num_decks=1, num_rounds=1, options=options)
|
||||
|
||||
game_id = logger.log_game_start(
|
||||
room_code="DETAIL",
|
||||
num_players=num_players,
|
||||
options=options
|
||||
)
|
||||
|
||||
# Initial flips
|
||||
print("\nInitial flips:")
|
||||
for player, profile in players_with_profiles:
|
||||
@ -622,7 +586,7 @@ def run_detailed_game(num_players: int = 4, options: Optional[GameOptions] = Non
|
||||
print(f" Discard: {discard_before.rank.value}")
|
||||
|
||||
# Run turn
|
||||
action = run_cpu_turn(game, current, profile, logger, game_id, stats)
|
||||
action = run_cpu_turn(game, current, profile, stats)
|
||||
|
||||
# Show result
|
||||
discard_after = game.discard_top()
|
||||
@ -635,8 +599,6 @@ def run_detailed_game(num_players: int = 4, options: Optional[GameOptions] = Non
|
||||
turn += 1
|
||||
|
||||
# Game over
|
||||
logger.log_game_end(game_id)
|
||||
|
||||
print("\n" + "=" * 50)
|
||||
print("FINAL SCORES")
|
||||
print("=" * 50)
|
||||
@ -648,8 +610,6 @@ def run_detailed_game(num_players: int = 4, options: Optional[GameOptions] = Non
|
||||
|
||||
winner = min(game.players, key=lambda p: p.total_score)
|
||||
print(f"\nWinner: {winner.name}!")
|
||||
|
||||
print(f"\nGame logged as: {game_id[:8]}...")
|
||||
print("Run: python game_analyzer.py game", game_id, winner.name)
|
||||
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user