Py-Hanabi/hanabi/solvers/greedy_solver.py

287 lines
12 KiB
Python
Executable file

#! /bin/python3
import collections
import sys
from enum import Enum
from hanabi.log_setup import logger
from typing import Optional
from hanabi import DeckCard, GameState, HanabiInstance
from hanabi.live.compress import link, decompress_deck
from hanabi.database.database import conn
class CardType(Enum):
Dispensable = -1
Trash = 0
Playable = 1
Critical = 2
DuplicateVisible = 3
UniqueVisible = 4
class CardState():
def __init__(self, card_type: CardType, card: DeckCard, weight=1):
self.card_type = card_type
self.card = card
self.weight = weight
def __repr__(self):
match self.card_type:
case CardType.Trash:
return "Trash ({})".format(self.card)
case CardType.Playable:
return "Playable ({}) with weight {}".format(self.card, self.weight)
case CardType.Critical:
return "Critical ({})".format(self.card)
case CardType.DuplicateVisible:
return "Useful (duplicate visible) ({}) with weight {}".format(self.card, self.weight)
case CardType.UniqueVisible:
return "Useful (unique visible) ({}) with weight {}".format(self.card, self.weight)
# TODO
def card_type(game_state, card):
played = game_state.stacks[card.suitIndex]
if card.rank <= played:
return CardType.Trash
elif card.rank == played + 1:
return CardType.Playable
elif card.rank == 5 or card in game_state.trash:
return CardType.Critical
else:
visible_cards = sum((game_state.hands[player] for player in range(game_state.num_players)), [])
if visible_cards.count(card) >= 2:
return CardType.DuplicateVisible
else:
return CardType.UniqueVisible
class WeightedCard:
def __init__(self, card, weight: Optional[int] = None):
self.card = card
self.weight = weight
def __repr__(self):
return "{} with weight {}".format(self.card, self.weight)
class HandState:
def __init__(self, player: int, game_state: GameState):
self.trash = []
self.playable = []
self.critical = []
self.dupes = []
self.uniques = []
for card in game_state.hands[player]:
match card_type(game_state, card):
case CardType.Trash:
self.trash.append(WeightedCard(card))
case CardType.Playable:
if card not in map(lambda c: c.card, self.playable):
self.playable.append(WeightedCard(card))
else:
self.trash.append(card)
case CardType.Critical:
self.critical.append(WeightedCard(card))
case CardType.UniqueVisible:
self.uniques.append(WeightedCard(card))
case CardType.DuplicateVisible:
copy = next((w for w in self.dupes if w.card == card), None)
if copy is not None:
self.dupes.remove(copy)
self.critical.append(copy)
self.trash.append(card)
else:
self.dupes.append(WeightedCard(card))
self.playable.sort(key=lambda c: c.card.rank)
self.dupes.sort(key=lambda c: c.card.rank)
self.uniques.sort(key=lambda c: c.card.rank)
if len(self.trash) > 0:
self.best_discard = self.trash[0]
self.discard_badness = 0
elif len(self.dupes) > 0:
self.best_discard = self.dupes[0]
self.discard_badness = 8 - game_state.num_players
elif len(self.uniques) > 0:
self.best_discard = self.uniques[-1]
self.discard_badness = 80 - 10 * self.best_discard.card.rank
elif len(self.playable) > 0:
self.best_discard = self.playable[-1]
self.discard_badness = 80 - 10 * self.best_discard.card.rank
else:
assert len(self.critical) > 0, "Programming error."
self.best_discard = self.critical[-1]
self.discard_badness = 600 - 100*self.best_discard.card.rank
def num_useful_cards(self):
return len(self.dupes) + len(self.uniques) + len(self.playable) + len(self.critical)
class CheatingStrategy:
def __init__(self, game_state: GameState):
self.game_state = game_state
def make_move(self):
hand_states = [HandState(player, self.game_state) for player in range(self.game_state.num_players)]
modified_pace = self.game_state.pace - sum(
1 for state in hand_states if len(state.trash) == self.game_state.hand_size
)
cur_hand = hand_states[self.game_state.turn]
print([state.__dict__ for state in hand_states])
print(self.game_state.pace)
exit(0)
class GreedyStrategy():
def __init__(self, game_state: GameState):
self.game_state = game_state
self.earliest_draw_times = []
for s in range(0, game_state.instance.num_suits):
self.earliest_draw_times.append([])
for r in range(1, 6):
self.earliest_draw_times[s].append(max(
game_state.deck.index(DeckCard(s, r)) - game_state.hand_size * game_state.num_players + 1,
0 if r == 1 else self.earliest_draw_times[s][r - 2]
))
# Currently, we do not add the time the 5 gets drawn to this, since this is rather a measurument on how
# bad a suit is in terms of having to hold on to other cards that are not playable *yet*
self.suit_badness = [sum(self.earliest_draw_times[s][:-1]) for s in range(0, game_state.num_suits)]
def make_move(self):
hand_states = [[CardState(card_type(self.game_state, card), card, None) for card in self.game_state.hands[p]]
for p in range(self.game_state.num_players)]
# find dupes in players hands, marke one card crit and the other one trash
p = False
for states in hand_states:
counter = collections.Counter(map(lambda state: state.card, states))
for card in counter:
if counter[card] >= 2:
dupes = (cstate for cstate in states if cstate.card == card)
first = next(dupes)
if first.card_type == CardType.Dispensable:
first.card_type = CardType.Critical
for dupe in dupes:
dupe.card_type = CardType.Trash
def hand_badness(states):
if any(state.card_type == CardType.Playable for state in states):
return 0
crits = [state for state in states if state.card_type == CardType.Critical]
crits_val = sum(map(lambda state: state.card.rank, crits))
if any(state.card_type == CardType.Playable for state in states):
return crits_val
def player_distance(f, t):
return ((t - f - 1) % self.game_state.num_players) + 1
for (player, states) in enumerate(hand_states):
for state in states:
if state.card_type == CardType.Playable:
copy_holders = set(self.game_state.holding_players(state.card))
copy_holders.remove(player)
connecting_holders = set(
self.game_state.holding_players(DeckCard(state.card.suitIndex, state.card.rank + 1)))
if len(copy_holders) == 0:
# card is unique, imortancy is based lexicographically on whether somebody has the conn. card and the rank
state.weight = (6 if len(connecting_holders) > 0 else 1) * (6 - state.card.rank)
else:
# copy is available somewhere else
if len(connecting_holders) == 0:
# card is not urgent
state.weight = 0.5 * (6 - state.card.rank)
else:
# there is a copy and there is a connecting card. check if they are out of order
turns_to_copy = min(map(lambda holder: player_distance(player, holder), copy_holders))
turns_to_conn = max(map(lambda holder: player_distance(player, holder), connecting_holders))
if turns_to_copy < turns_to_conn:
# our copy is not neccessary for connecting card to be able to play
state.weight = 0.5 * (6 - state.card.rank)
else:
# our copy is important, scale it little less than if it were unique
state.weight = 4 * (6 - state.card.rank)
elif state.card_type == CardType.Dispensable:
try:
# TODO: consider duplicate in hand
copy_holders = list(self.game_state.holding_players(state.card))
copy_holders.remove(player)
nextCopy = self.game_state.deck[self.game_state.progress:].index(card)
except:
nextCopy = 1
# state.weight = self.suit_badness[state.card.suitIndex] * nextCopy + 2 * (5 - state.card.rank)
state.weight = nextCopy + 2 * (5 - state.card.rank)
cur_hand = hand_states[self.game_state.turn]
plays = [cstate for cstate in cur_hand if cstate.card_type == CardType.Playable]
trash = next((cstate for cstate in cur_hand if cstate.card_type == CardType.Trash), None)
# actual decision on what to do
if len(plays) > 0:
play = max(plays, key=lambda s: s.weight)
self.game_state.play(play.card.deck_index)
elif self.game_state.clues == 8:
self.game_state.clue()
elif trash is not None:
self.game_state.discard(trash.card.deck_index)
elif self.game_state.clues == 0:
dispensable = [cstate for cstate in cur_hand if cstate.card_type == CardType.Dispensable]
if len(dispensable) == 0:
self.game_state.in_lost_state = True
# raise ValueError("Lost critical card")
else:
discard = min(dispensable, key=lambda s: s.weight)
self.game_state.discard(discard.card.deck_index)
else:
self.game_state.clue()
def run_deck(instance: HanabiInstance) -> GameState:
gs = GameState(instance)
strat = CheatingStrategy(gs)
while not gs.is_over():
strat.make_move()
return gs
def run_samples(num_players, sample_size):
logger.info("Running {} test games on {} players using greedy strategy.".format(sample_size, num_players))
won = 0
lost = 0
cur = conn.cursor()
cur.execute(
"SELECT seed, num_players, deck, variant_id "
"FROM seeds WHERE variant_id = 0 AND num_players = (%s)"
"ORDER BY seed DESC LIMIT (%s)",
(num_players, sample_size))
for r in cur:
seed, num_players, deck_str, var_id = r
deck = decompress_deck(deck_str)
instance = HanabiInstance(deck, num_players)
final_game_state = run_deck(instance)
if final_game_state.score != instance.max_score:
logger.verbose(
"Greedy strategy lost {}-player seed {:10} {}:\n{}"
.format(num_players, seed, str(deck), link(final_game_state))
)
lost += 1
else:
won += 1
print("won: {:4}, lost: {:4}".format(won, lost), end="\r")
logger.info("Won {} ({}%) and lost {} ({}%) from sample of {} test games using greedy strategy.".format(
won, round(100 * won / sample_size, 2), lost, round(100 * lost / sample_size, 2), sample_size
))
if __name__ == "__main__":
for p in range(2, 6):
run_samples(p, int(sys.argv[1]))
print()