def update_board(self, game_state: GameState):
assert isinstance(game_state, SpookState)
super().update_board(game_state)
valid_moves = game_state.get_valid_moves()
volume = game_state.calculate_volume()
is_pass_valid = bool(valid_moves[volume])
self.ui.pass_button.setVisible(is_pass_valid)
def analyse(self, board: GameState) -> typing.Tuple[float, np.ndarray]:
policy = self.get_policy(board)
player = board.get_active_player()
if board.is_win(player):
value = 1.0
elif board.is_win(-player):
value = -1.0
else:
value = 0.0
return value, policy
def analyse(self, board: GameState) -> typing.Tuple[float, np.ndarray]:
if board.is_ended():
return self.analyse_end_game(board)
outputs = self.model.predict(board.get_spaces().reshape(
(1,
self.board_height,
self.board_width,
1)))
policy = outputs[0, :-1]
value = outputs[0, -1]
return value, policy
def update_board(self, game_state: GameState):
assert isinstance(game_state, SpaijiState)
valid_moves = game_state.get_valid_moves()
volume = game_state.calculate_volume()
is_black_valid = valid_moves[:volume].any()
is_white_valid = valid_moves[volume:].any()
visible_move_types = []
if is_black_valid:
visible_move_types.append(MoveType.BLACK)
if is_white_valid:
visible_move_types.append(MoveType.WHITE)
self.visible_move_types = visible_move_types
super().update_board(game_state)
def get_policy(self, board: GameState):
valid_moves = board.get_valid_moves()
if not valid_moves.any():
valid_moves = (valid_moves == 0)
raw_policy = np.multiply(valid_moves, 0.9**np.arange(len(valid_moves)))
policy = raw_policy / raw_policy.sum()
return policy
def get_move_probabilities(
self,
game_state: GameState,
limit: int = 10
) -> typing.List[typing.Tuple[str, float, int, float]]:
""" Report the probability that each move is the best choice.
:param game_state: the starting position
:param limit: the maximum number of moves to report
:return: [(move_display, probability, value_count, avg_value)], where
value_count is the number of times the value was probed from the move,
and avg_value is the average value from all those probes.
"""
self.find_node(game_state)
children = self.current_node.find_all_children()
temperature = 1.0
probabilities = self.current_node.rank_children(children, temperature)
value_counts = [child.value_count for child in children]
ranked_children = sorted(zip(value_counts, probabilities, children),
key=itemgetter(0),
reverse=True)
top_children = ranked_children[:limit]
child_node: SearchNode
top_moves = [(game_state.display_move(child_node.move), probability,
value_count, child_node.average_value)
for value_count, probability, child_node in top_children
if child_node.move is not None]
return top_moves
def analyse_end_game(self,
board: GameState) -> typing.Tuple[float, np.ndarray]:
""" Calculate the value based on the winner.
:param board: state of a game, that must be in an ended game
:return: 1 if the now active player won the game, -1 for a loss, and
0 for a draw, plus an evenly distributed policy (probably irrelevant)
"""
winner = board.get_winner()
active_player = board.get_active_player()
if winner == board.NO_PLAYER:
value = 0
elif winner == active_player:
value = 1
else:
value = -1
return value, self.create_even_policy(board)
def update_board(self, game_state: GameState):
assert isinstance(game_state, SpireState)
move_types = game_state.get_valid_colours()
self.visible_move_types = move_types
if self.selected_move_type not in move_types:
self.selected_move_type = move_types[0]
super().update_board(game_state)
def get_policy(self, board: GameState):
valid_moves = board.get_valid_moves()
if valid_moves.any():
first_valid = np.nonzero(valid_moves)[0][0]
else:
first_valid = 0
policy = np.zeros_like(valid_moves)
policy[first_valid] = 1.0
return policy
def choose_move(self, game_state: GameState) -> int:
""" Choose a move for the given board.
:param game_state: the current state of the game.
:return: the chosen move's index in the list of valid moves.
"""
self.search_manager.search(game_state, self.iteration_count)
if game_state.get_move_count() < 15:
return self.search_manager.choose_weighted_move()
return self.search_manager.get_best_move()
def create_even_policy(board: GameState) -> np.ndarray:
""" Create an evenly distributed policy across valid moves.
If there are no valid moves, distribute across all entries.
"""
child_valid_flags = board.get_valid_moves()
valid_count = child_valid_flags.sum()
if valid_count:
child_predictions = child_valid_flags / valid_count
else:
child_predictions = (child_valid_flags +
1) / child_valid_flags.size
return child_predictions
def __init__(self, start_state: GameState):
if not isinstance(start_state, GridGameState):
raise ValueError(f'{start_state.__class__} is not a subclass of GridGameState.')
super().__init__()
# start_state params
self.board_height = start_state.board_height
self.board_width = start_state.board_width
self.action_size = len(start_state.get_valid_moves())
self.epochs_completed = 0
self.epochs_to_train = 100
self.start_state = start_state
args = Namespace(lr=0.001,
dropout=0.3,
epochs=10,
batch_size=64,
num_channels=512)
self.checkpoint_name = 'random weights'
self.args = args
num_channels = 512
kernel_size = [3, 3]
dropout = 0.3
model = Sequential()
# regularizer = regularizers.l2(0.00006)
regularizer = regularizers.l2(0.0001)
model.add(Conv2D(num_channels,
kernel_size,
padding='same',
activation='relu',
input_shape=(self.board_height, self.board_width, 1),
activity_regularizer=regularizer))
model.add(Conv2D(num_channels,
kernel_size,
padding='same',
activation='relu',
activity_regularizer=regularizer))
model.add(Conv2D(num_channels,
kernel_size,
activation='relu',
activity_regularizer=regularizer))
model.add(Conv2D(num_channels,
kernel_size,
activation='relu',
activity_regularizer=regularizer))
model.add(Dropout(dropout))
model.add(Dropout(dropout))
model.add(Flatten())
model.add(Dense(self.action_size + 1))
model.compile('adam', 'mean_squared_error')
self.model = model
def analyse_move(self, game_state: GameState, analysing_player: int,
move_probabilities: typing.List[typing.Tuple[str, float,
int,
float]]):
""" Record analysis of the choices for a move.
:param game_state: the state before the move
:param analysing_player: the player doing the analysis; if both players
report analysis, the active player's will be kept
:param move_probabilities: the detailed analysis of best valid moves
[(choice, probability, count, value)] where choice is the move
display, probability is the recommended probability of choosing that
move, count is the number of times the move or a descendant was
analysed during the search, and value is the estimated value of the
move, where 1 is 100% wins for the active player and -1 is 100%
losses.
"""
for item in reversed(self.items):
if item.game_state == game_state:
break
else:
raise ValueError('Board not found in log.')
active_player = game_state.get_active_player()
if item.choices and active_player != analysing_player:
return
item.choices = move_probabilities
q = Decimal('0.1')
for i, (choice, probability, count,
value) in enumerate(move_probabilities, 1):
dec_value = Decimal(value).quantize(q)
if choice == item.move_text:
item.comment = f'{dec_value}'
if i != 1:
item.comment += f', choice {i}'
break
else:
# Didn't find chosen move in analysed moves, probably a poor choice.
item.comment = '?'
def record_move(self, game_state: GameState, move: int):
self.step += 1
player = game_state.display_player(game_state.get_active_player())
move_text = game_state.display_move(move)
self.items.append(LogItem(self.step, player, move_text, game_state))
def update_board(self, game_state: GameState):
assert isinstance(game_state, ShibumiGameState)
self.current_state = game_state
valid_moves = game_state.get_valid_moves()
is_ended = game_state.is_ended()
pixmaps = {
game_state.BLACK: self.black_scaled,
game_state.WHITE: self.white_scaled,
game_state.RED: self.red_scaled,
game_state.NO_PLAYER: self.black_scaled
}
if self.show_move_types:
move_type = self.selected_move_type
else:
move_type = MoveType.BLACK
for level, item_level in zip(game_state.get_levels(),
self.item_levels):
for row_pieces, row_items in zip(level, item_level):
piece_item: GraphicsShibumiPieceItem
for piece, piece_item in zip(row_pieces, row_items):
move_index = game_state.get_index(piece_item.height,
piece_item.row,
piece_item.column,
move_type)
is_valid = valid_moves[move_index] and not is_ended
piece_item.setVisible(
bool(piece != game_state.NO_PLAYER or is_valid))
piece_item.hover_listener = self if is_valid else None
pixmap = pixmaps[piece]
piece_item.setPixmap(pixmap)
piece_item.setOpacity(0.001 if piece ==
game_state.NO_PLAYER else 1)
if piece_item == self.hovered_piece:
if is_valid:
piece_item.setOpacity(0.5)
else:
self.hovered_piece = None
displayed_player: typing.Optional[int] = None
if is_ended:
if game_state.is_win(game_state.WHITE):
self.update_move_text('wins')
displayed_player = game_state.WHITE
elif game_state.is_win(game_state.BLACK):
self.update_move_text('wins')
displayed_player = game_state.BLACK
elif game_state.is_win(game_state.RED):
self.update_move_text('wins')
displayed_player = game_state.RED
else:
self.update_move_text('draw')
else:
displayed_player = game_state.get_active_player()
self.update_move_text(self.choose_active_text())
if displayed_player == game_state.WHITE:
self.ui.player_pixmap.setPixmap(self.white_pixmap)
elif displayed_player == game_state.BLACK:
self.ui.player_pixmap.setPixmap(self.black_pixmap)
elif displayed_player == game_state.RED:
self.ui.player_pixmap.setPixmap(self.red_pixmap)
if displayed_player is not None:
displayed_player = int(displayed_player)
self.ui.player_pixmap.setVisible(
displayed_player != game_state.NO_PLAYER)
if self.show_counts:
self.update_count_text()
def update_board(self, game_state: GameState):
player = game_state.get_active_player()
player_move = MoveType(player)
self.visible_move_types = (player_move, MoveType.RED)
super().update_board(game_state)
def on_game_ended(self, game_state: GameState):
if (self.is_history_dirty or self.display is None
or self.ui.searches_lock1.isChecked()):
return
db_session = self.db_session
if db_session is None:
return
game_record = GameRecord.find_or_create(db_session, game_state)
game_end_time = datetime.now()
game_duration = game_end_time - self.game_start_time
match_record = MatchRecord(game=game_record,
start_time=self.game_start_time,
total_seconds=round(
game_duration.total_seconds()),
move_count=game_state.get_move_count())
db_session.add(match_record)
winner = game_state.get_winner()
mcts_player: typing.Optional[MctsPlayer]
for player_number in game_state.get_players():
mcts_player = self.display.get_player(player_number)
if mcts_player is None:
player_record = db_session.query(PlayerRecord).filter_by(
type=PlayerRecord.HUMAN_TYPE).one_or_none()
if player_record is None:
player_record = PlayerRecord(type=PlayerRecord.HUMAN_TYPE)
db_session.add(player_record)
else:
player_record = db_session.query(PlayerRecord).filter_by(
type=PlayerRecord.PLAYOUT_TYPE,
iterations=mcts_player.iteration_count).one_or_none()
if player_record is None:
player_record = PlayerRecord(
type=PlayerRecord.PLAYOUT_TYPE,
iterations=mcts_player.iteration_count)
db_session.add(player_record)
if player_number == winner:
result = 1
elif winner == game_state.NO_PLAYER:
result = 0
else:
result = -1
match_player = MatchPlayerRecord(match=match_record,
player=player_record,
player_number=player_number,
result=result)
db_session.add(match_player)
db_session.commit()
try:
mcts_player, = self.display.mcts_players
except ValueError:
# Didn't have exactly one MCTS player
return
assert mcts_player is not None
winning_player = game_state.get_winner()
if winning_player == mcts_player.player_number:
score = -1
elif winning_player == GameState.NO_PLAYER:
score = 0
else:
score = 1
settings = get_settings(self.start_state)
strength_adjuster = StrengthAdjuster(
strength=mcts_player.iteration_count,
game_count=settings.value('game_count', 0, int),
last_score=settings.value('last_score', 0, int),
streak_length=settings.value('streak_length', 1, int))
strength_adjuster.record_score(score)
settings.setValue('searches', strength_adjuster.strength)
settings.setValue('game_count', strength_adjuster.game_count)
settings.setValue('last_score', strength_adjuster.last_score)
settings.setValue('streak_length', strength_adjuster.streak_length)
