Hanabi/Py-Hanabi
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

rework analysis of upper bounds: compute all bounds now, insert into DB properly

Browse source
This commit is contained in:
Maximilian Keßler 2023-07-08 09:48:22 +02:00
parent 91f3c73eb3
commit 29cae8f139
Signed by: max
GPG key ID: BCC5A619923C0BA5
1 changed files with 129 additions and 134 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

263
hanabi/solvers/deck_analyzer.py
View file

@ -1,7 +1,8 @@
from hanabi.live import compress
from enum import Enum
from typing import List
from hanabi import database
from hanabi import logger
from hanabi import hanab_game
from hanabi.live import compress
@ -9,107 +10,85 @@ from hanabi.live import compress
class InfeasibilityType(Enum):
OutOfPace = 0 # idx denotes index of last card drawn before being forced to reduce pace, value denotes how bad pace is
OutOfHandSize = 1 # idx denotes index of last card drawn before being forced to discard a crit
NotTrivial = 2
CritAtBottom = 3
class InfeasibilityReason():
def __init__(self, infeasibility_type, idx, value=None):
class InfeasibilityReason:
def __init__(self, infeasibility_type: InfeasibilityType, score_upper_bound, value=None):
self.type = infeasibility_type
self.index = idx
self.score_upper_bound = score_upper_bound
self.value = value
def __repr__(self):
match self.type:
case InfeasibilityType.OutOfPace:
return "Deck runs out of pace ({}) after drawing card {}".format(self.value, self.index)
return "Upper bound {}, since deck runs out of pace after drawing card {}".format(self.score_upper_bound, self.value)
case InfeasibilityType.OutOfHandSize:
return "Deck runs out of hand size after drawing card {}".format(self.index)
return "Upper bound {}, since deck runs out of hand size after drawing card {}".format(self.score_upper_bound, self.value)
case InfeasibilityType.CritAtBottom:
return "Deck has crit non-5 at bottom (index {})".format(self.index)
return "Upper bound {}, sicne deck has critical non-5 at bottom".format(self.score_upper_bound)
def analyze_suit(occurrences):
# denotes the indexes of copies we can use wlog
picks = {
1: 0,
**{r: None for r in range(2, 5)},
5: 0
}
def analyze(instance: hanab_game.HanabiInstance, only_find_first=False) -> List[InfeasibilityReason]:
"""
Checks instance for the following (easy) certificates for unfeasibility
- There is a critical non-5 at the bottom
- We necessarily run out of pace when playing this deck:
At some point, among all drawn cards, there are too few playable ones so that the next discard
reduces pace to a negative amount
- We run out of hand size when playing this deck:
At some point, there are too many critical cards (that also could not have been played) for the players
to hold collectively
:param instance: Instance to be analyzed
:param only_find_first: If true, we immediately return when finding the first infeasibility reason and don't
check for further ones. Might be slightly faster on some instances, especially dark ones.
:return: List with all reasons found. Empty if none is found.
In particular, if return value is not the empty list, the analyzed instance is unfeasible
"""
reasons = []
# denotes the intervals when cards will be played wlog
play_times = {
1: [occurrences[1][0]],
**{r: None for _ in range(instance.num_suits)
for r in range(2, 6)
}
}
# check for critical non-fives at bottom of the deck
bottom_card = instance.deck[-1]
if bottom_card.rank != 5 and bottom_card.suitIndex in instance.dark_suits:
reasons.append(InfeasibilityReason(
InfeasibilityType.CritAtBottom,
instance.max_score - 5 + bottom_card.rank,
instance.deck_size - 1
))
if only_find_first:
return reasons
print("occurrences are: {}".format(occurrences))
for rank in range(2, 6):
# general analysis
earliest_play = max(min(play_times[rank - 1]), min(occurrences[rank]))
latest_play = max(*play_times[rank - 1], *occurrences[rank])
play_times[rank] = [earliest_play, latest_play]
# check a few extra cases regarding the picks when the rank is not 5
if rank != 5:
# check if we can just play the first copy
if max(play_times[rank - 1]) < min(occurrences[rank]):
picks[rank] = 0
play_times[rank] = [min(occurrences[rank])]
continue
# check if the second copy is not worse than the first when it comes,
# because we either have to wait for smaller cards anyway
# or the next card is not there anyway
if max(occurrences[rank]) < max(earliest_play, min(occurrences[rank + 1])):
picks[rank] = 1
return picks, play_times
def analyze_card_usage(instance: hanab_game.HanabiInstance):
storage_size = instance.num_players * instance.hand_size
for suit in range(instance.num_suits):
print("analysing suit {}: {}".format(
suit,
hanab_game.pp_deck((c for c in instance.deck if c.suitIndex == suit))
)
)
occurrences = {
rank: [max(0, i - storage_size + 1) for (i, card) in enumerate(instance.deck) if
card == hanab_game.DeckCard(suit, rank)]
for rank in range(1, 6)
}
picks, play_times = analyze_suit(occurrences)
print("did analysis:")
print("play times: ", play_times)
print("picks: ", picks)
print()
def analyze(instance: hanab_game.HanabiInstance, find_non_trivial=False) -> InfeasibilityReason | None:
if instance.deck[-1].rank != 5 and instance.deck[-1].suitIndex + instance.num_dark_suits >= instance.num_suits:
return InfeasibilityReason(InfeasibilityType.CritAtBottom, instance.deck_size - 1)
# we will sweep through the deck and pretend that we instantly play all cards
# as soon as we have them (and recurse this)
# we will sweep through the deck and pretend that
# - we keep all non-trash cards in our hands
# - we instantly play all playable cards as soon as we have them
# - we recurse on this instant-play
#
# For example, we assume that once we draw r2, we check if we can play r2.
# If yes, then we also check if we drew r3 earlier and so on.
# If not, then we keep r2 in our hands
#
# In total, this is equivalent to assuming that we infinitely many clues
# and infinite storage space in our hands (which is of course not true),
# but even in this setting, some games are infeasible due to pace issues
# that we can detect
#
# A small refinement is to pretend that we only have infinite storage for non-crit cards,
# for crit-cards, the usual hand card limit applies.
# This allows us to detect some seeds where there are simply too many unplayable cards to hold at some point
# that also can't be discarded
# this allows us to detect standard pace issue arguments
stacks = [0] * instance.num_suits
# we will ensure that stored_crits is a subset of stored_cards
stored_cards = set()
stored_crits = set()
min_forced_pace = 100
worst_index = 0
min_forced_pace = instance.initial_pace
worst_pace_index = 0
ret = None
max_forced_crit_discard = 0
worst_crit_index = 0
for (i, card) in enumerate(instance.deck):
if card.rank == stacks[card.suitIndex] + 1:
@ -133,68 +112,84 @@ def analyze(instance: hanab_game.HanabiInstance, find_non_trivial=False) -> Infe
stored_crits.add(card)
stored_cards.add(card)
# check for out of handsize:
if len(stored_crits) == instance.num_players * instance.hand_size:
return InfeasibilityReason(InfeasibilityType.OutOfHandSize, i)
if find_non_trivial and len(stored_cards) == instance.num_players * instance.hand_size:
ret = InfeasibilityReason(InfeasibilityType.NotTrivial, i)
# check for out of handsize (this number can be negative, in which case nothing applies)
# Note the +1 at the end, which is there because we have to discard next,
# so even if we currently have as many crits as we can hold, we have to discard one
num_forced_crit_discards = len(stored_crits) - instance.num_players * instance.hand_size + 1
if len(stored_crits) - instance.num_players * instance.hand_size > max_forced_crit_discard:
worst_crit_index = i
max_forced_crit_discard = num_forced_crit_discards
if only_find_first:
reasons.append(InfeasibilityReason(
InfeasibilityType.OutOfPace,
instance.max_score + min_forced_pace,
worst_pace_index
))
return reasons
# the last - 1 is there because we have to discard 'next', causing a further draw
max_remaining_plays = (instance.deck_size - i - 1) + instance.num_players - 1
needed_plays = 5 * instance.num_suits - sum(stacks)
missing = max_remaining_plays - needed_plays
if missing < min_forced_pace:
# print("update to {}: {}".format(i, missing))
min_forced_pace = missing
worst_index = i
needed_plays = instance.max_score - sum(stacks)
cur_pace = max_remaining_plays - needed_plays
if cur_pace < min(0, min_forced_pace):
min_forced_pace = cur_pace
worst_pace_index = i
if only_find_first:
reasons.append(InfeasibilityReason(
InfeasibilityType.OutOfPace,
instance.max_score + min_forced_pace,
worst_pace_index
))
return reasons
# check that we correctly walked through the deck
assert (len(stored_cards) == 0)
assert (len(stored_crits) == 0)
assert (sum(stacks) == 5 * instance.num_suits)
assert (sum(stacks) == instance.max_score)
if max_forced_crit_discard > 0:
reasons.append(
InfeasibilityReason(
InfeasibilityType.OutOfHandSize,
instance.max_score - max_forced_crit_discard,
worst_crit_index
)
)
if min_forced_pace < 0:
return InfeasibilityReason(InfeasibilityType.OutOfPace, worst_index, min_forced_pace)
elif ret is not None:
return ret
else:
return None
reasons.append(InfeasibilityReason(
InfeasibilityType.OutOfPace,
instance.max_score + min_forced_pace,
worst_pace_index
))
return reasons
def run_on_database():
cur = database.conn.cursor()
cur2 = database.conn.cursor()
for num_p in range(2, 6):
cur.execute(
"SELECT seed, num_players, deck from seeds where variant_id = 0 and num_players = (%s) order by seed asc",
(num_p,))
res = cur.fetchall()
hand = 0
pace = 0
non_trivial = 0
d = None
print("Checking {} {}-player seeds from database".format(len(res), num_p))
for (seed, num_players, deck) in res:
deck = compress.decompress_deck(deck)
a = analyze(hanab_game.HanabiInstance(deck, num_players), True)
if type(a) == InfeasibilityReason:
if a.type == InfeasibilityType.OutOfHandSize:
# print("Seed {} infeasible: {}\n{}".format(seed, a, deck))
hand += 1
elif a.type == InfeasibilityType.OutOfPace:
pace += 1
elif a.type == InfeasibilityType.NotTrivial:
non_trivial += 1
d = seed, deck
print("Found {} seeds running out of hand size, {} running out of pace and {} that are not trivial".format(hand,
pace,
non_trivial))
if d is not None:
print("example non-trivial deck (seed {}): [{}]".format(
d[0],
", ".join(c.colorize() for c in d[1])
))
print()
def run_on_database(variant_id):
database.cur.execute(
"SELECT seed, num_players, deck FROM seeds WHERE variant_id = (%s) ORDER BY (num_players, seed)",
(variant_id,)
)
res = database.cur.fetchall()
logger.info("Checking {} seeds of variant {} for infeasibility".format(len(res), variant_id))
for (seed, num_players, deck_str) in res:
deck = compress.decompress_deck(deck_str)
reasons = analyze(hanab_game.HanabiInstance(deck, num_players))
if reasons:
print("found infeasible seed {}: {}".format(seed, reasons))
else:
print("found nothing for seed {}".format(seed))
for reason in reasons:
database.cur.execute(
"INSERT INTO score_upper_bounds (seed, score_upper_bound, reason) "
"VALUES (%s,%s,%s) "
"ON CONFLICT (seed, reason) DO UPDATE "
"SET score_upper_bound = EXCLUDED.score_upper_bound",
(seed, reason.score_upper_bound, reason.type.value)
)
database.cur.execute(
"UPDATE seeds SET feasible = (%s) WHERE seed = (%s)",
(False, seed)
)
database.conn.commit()