def test_static_env():
from cchess_alphazero.environment.env import CChessEnv
import cchess_alphazero.environment.static_env as senv
from cchess_alphazero.environment.static_env import INIT_STATE
from cchess_alphazero.environment.lookup_tables import flip_move
env = CChessEnv()
env.reset()
print("env: " + env.observation)
print("senv: " + INIT_STATE)
state = INIT_STATE
env.step('0001')
state = senv.step(state, '0001')
print(senv.evaluate(state))
print("env: " + env.observation)
print("senv: " + state)
env.step('7770')
state = senv.step(state, flip_move('7770'))
print(senv.evaluate(state))
print("env: " + env.observation)
print("senv: " + state)
env.render()
board = senv.state_to_board(state)
for i in range(9, -1, -1):
print(board[i])
print("env: ")
print(env.input_planes()[0+7:3+7])
print("senv: ")
print(senv.state_to_planes(state)[0+7:3+7])
print(f"env: {env.board.legal_moves()}" )
print(f"senv: {senv.get_legal_moves(state)}")
print(set(env.board.legal_moves()) == set(senv.get_legal_moves(state)))
def load_game(self, init, move_list, winner, idx, title, url):
turns = 0
env = CChessEnv(self.config).reset(init)
red_moves = []
black_moves = []
moves = [
move_list[i:i + 4] for i in range(len(move_list)) if i % 4 == 0
]
for move in moves:
action = senv.parse_onegreen_move(move)
try:
if turns % 2 == 0:
red_moves.append([
env.observation,
self.build_policy(action, flip=False)
])
else:
black_moves.append([
env.observation,
self.build_policy(action, flip=True)
])
env.step(action)
except:
logger.error(
f"Invalid Action: idx = {idx}, action = {action}, turns = {turns}, moves = {moves}, "
f"winner = {winner}, init = {init}, title: {title}, url: {url}"
)
return
turns += 1
if winner == Winner.red:
red_win = 1
elif winner == Winner.black:
red_win = -1
else:
red_win = senv.evaluate(env.get_state())
if not env.red_to_move:
red_win = -red_win
for move in red_moves:
move += [red_win]
for move in black_moves:
move += [-red_win]
data = []
for i in range(len(red_moves)):
data.append(red_moves[i])
if i < len(black_moves):
data.append(black_moves[i])
self.buffer += data
return red_win
def load_game(self, init, move_list, winner, idx, title, url):
turns = 0
env = CChessEnv(self.config).reset(init)
red_moves = []
black_moves = []
moves = [move_list[i:i+4] for i in range(len(move_list)) if i % 4 == 0]
for move in moves:
action = senv.parse_onegreen_move(move)
try:
if turns % 2 == 0:
red_moves.append([env.observation, self.build_policy(action, flip=False)])
else:
black_moves.append([env.observation, self.build_policy(action, flip=True)])
env.step(action)
except:
logger.error(f"Invalid Action: idx = {idx}, action = {action}, turns = {turns}, moves = {moves}, "
f"winner = {winner}, init = {init}, title: {title}, url: {url}")
return
turns += 1
if winner == Winner.red:
red_win = 1
elif winner == Winner.black:
red_win = -1
else:
red_win = senv.evaluate(env.get_state())
if not env.red_to_move:
red_win = -red_win
for move in red_moves:
move += [red_win]
for move in black_moves:
move += [-red_win]
data = []
for i in range(len(red_moves)):
data.append(red_moves[i])
if i < len(black_moves):
data.append(black_moves[i])
self.buffer += data
return red_win
def start_game(self, idx, search_tree):
pipes = self.cur_pipes.pop()
if not self.config.play.share_mtcs_info_in_self_play or \
idx % self.config.play.reset_mtcs_info_per_game == 0:
search_tree = defaultdict(VisitState)
if random() > self.config.play.enable_resign_rate:
enable_resign = True
else:
enable_resign = False
self.player = CChessPlayer(self.config,
search_tree=search_tree,
pipes=pipes,
enable_resign=enable_resign,
debugging=False)
state = senv.INIT_STATE
history = [state]
policys = []
value = 0
turns = 0 # even == red; odd == black
game_over = False
final_move = None
while not game_over:
no_act = None
if state in history[:-1]:
no_act = []
for i in range(len(history) - 1):
if history[i] == state:
no_act.append(history[i + 1])
start_time = time()
action, policy = self.player.action(state, turns, no_act)
end_time = time()
if action is None:
logger.debug(f"{turns % 2} (0 = red; 1 = black) has resigned!")
value = -1
break
# logger.debug(f"Process{self.pid} Playing: {turns % 2}, action: {action}, time: {(end_time - start_time):.1f}s")
# for move, action_state in self.player.search_results.items():
# if action_state[0] >= 20:
# logger.info(f"move: {move}, prob: {action_state[0]}, Q_value: {action_state[1]:.2f}, Prior: {action_state[2]:.3f}")
# self.player.search_results = {}
history.append(action)
policys.append(policy)
state = senv.step(state, action)
turns += 1
history.append(state)
if turns / 2 >= self.config.play.max_game_length:
game_over = True
value = senv.evaluate(state)
else:
game_over, value, final_move = senv.done(state)
if final_move:
policy = self.build_policy(final_move, False)
history.append(final_move)
policys.append(policy)
state = senv.step(state, final_move)
history.append(state)
self.player.close()
if turns % 2 == 1: # balck turn
value = -value
v = value
if v == 0:
if random() > 0.5:
store = True
else:
store = False
else:
store = True
if store:
data = []
for i in range(turns):
k = i * 2
data.append([history[k], policys[i], value])
value = -value
self.save_play_data(idx, data)
self.cur_pipes.append(pipes)
self.remove_play_data()
return v, turns, state, search_tree, store
def start_game(self, idx, search_tree):
pipes = self.cur_pipes.pop()
if not self.config.play.share_mtcs_info_in_self_play or \
idx % self.config.play.reset_mtcs_info_per_game == 0:
search_tree = defaultdict(VisitState)
if random() > self.config.play.enable_resign_rate:
enable_resign = True
else:
enable_resign = False
self.player = CChessPlayer(self.config, search_tree=search_tree, pipes=pipes, enable_resign=enable_resign, debugging=False)
state = senv.INIT_STATE
history = [state]
policys = []
value = 0
turns = 0 # even == red; odd == black
game_over = False
final_move = None
while not game_over:
no_act = None
if state in history[:-1]:
no_act = []
for i in range(len(history) - 1):
if history[i] == state:
no_act.append(history[i + 1])
start_time = time()
action, policy = self.player.action(state, turns, no_act)
end_time = time()
if action is None:
logger.debug(f"{turns % 2} (0 = red; 1 = black) has resigned!")
value = -1
break
# logger.debug(f"Process{self.pid} Playing: {turns % 2}, action: {action}, time: {(end_time - start_time):.1f}s")
# for move, action_state in self.player.search_results.items():
# if action_state[0] >= 20:
# logger.info(f"move: {move}, prob: {action_state[0]}, Q_value: {action_state[1]:.2f}, Prior: {action_state[2]:.3f}")
# self.player.search_results = {}
history.append(action)
policys.append(policy)
state = senv.step(state, action)
turns += 1
history.append(state)
if turns / 2 >= self.config.play.max_game_length:
game_over = True
value = senv.evaluate(state)
else:
game_over, value, final_move = senv.done(state)
if final_move:
policy = self.build_policy(final_move, False)
history.append(final_move)
policys.append(policy)
state = senv.step(state, final_move)
history.append(state)
self.player.close()
if turns % 2 == 1: # balck turn
value = -value
v = value
if v == 0:
if random() > 0.5:
store = True
else:
store = False
else:
store = True
if store:
data = []
for i in range(turns):
k = i * 2
data.append([history[k], policys[i], value])
value = -value
self.save_play_data(idx, data)
self.cur_pipes.append(pipes)
self.remove_play_data()
return v, turns, state, search_tree, store
def self_play_buffer(config, cur) -> (tuple, list):
pipe = cur.pop() # borrow
if random() > config.play.enable_resign_rate:
enable_resign = True
else:
enable_resign = False
player = CChessPlayer(config,
search_tree=defaultdict(VisitState),
pipes=pipe,
enable_resign=enable_resign,
debugging=False)
state = senv.INIT_STATE
history = [state]
policys = []
value = 0
turns = 0
game_over = False
final_move = None
while not game_over:
no_act = None
if state in history[:-1]:
no_act = []
for i in range(len(history) - 1):
if history[i] == state:
no_act.append(history[i + 1])
start_time = time()
action, policy = player.action(state, turns, no_act)
end_time = time()
if action is None:
logger.debug(f"{turns % 2} (0 = red; 1 = black) has resigned!")
value = -1
break
# logger.debug(f"Playing: {turns % 2}, action: {action}, time: {(end_time - start_time):.1f}s")
policys.append(policy)
history.append(action)
state = senv.step(state, action)
turns += 1
history.append(state)
if turns / 2 >= config.play.max_game_length:
game_over = True
value = senv.evaluate(state)
else:
game_over, value, final_move = senv.done(state)
if final_move:
policy = build_policy(final_move, False)
history.append(final_move)
policys.append(policy)
state = senv.step(state, final_move)
history.append(state)
player.close()
if turns % 2 == 1: # balck turn
value = -value
v = value
data = []
for i in range(turns):
k = i * 2
data.append([history[k], policys[i], value])
value = -value
cur.append(pipe)
return (turns, v), data
def start_game(self, idx, search_tree):
pipes = self.cur_pipes.pop()
if not self.config.play.share_mtcs_info_in_self_play or \
idx % self.config.play.reset_mtcs_info_per_game == 0:
search_tree = defaultdict(VisitState)
if random() > self.config.play.enable_resign_rate:
enable_resign = True
else:
enable_resign = False
self.player = CChessPlayer(self.config, search_tree=search_tree, pipes=pipes, enable_resign=enable_resign, debugging=False)
state = senv.INIT_STATE
history = [state]
policys = []
value = 0
turns = 0 # even == red; odd == black
game_over = False
is_alpha_red = True if idx % 2 == 0 else False
final_move = None
while not game_over:
if (is_alpha_red and turns % 2 == 0) or (not is_alpha_red and turns % 2 == 1):
no_act = None
if state in history[:-1]:
no_act = []
for i in range(len(history) - 1):
if history[i] == state:
no_act.append(history[i + 1])
action, policy = self.player.action(state, turns, no_act)
if action is None:
logger.debug(f"{turns % 2} (0 = red; 1 = black) has resigned!")
value = -1
break
else:
fen = senv.state_to_fen(state, turns)
action = self.get_ucci_move(fen)
if action is None:
logger.debug(f"{turns % 2} (0 = red; 1 = black) has resigned!")
value = -1
break
if turns % 2 == 1:
action = flip_move(action)
try:
policy = self.build_policy(action, False)
except Exception as e:
logger.error(f"Build policy error {e}, action = {action}, state = {state}, fen = {fen}")
value = 0
break
history.append(action)
policys.append(policy)
state = senv.step(state, action)
turns += 1
history.append(state)
if turns / 2 >= self.config.play.max_game_length:
game_over = True
value = senv.evaluate(state)
else:
game_over, value, final_move = senv.done(state)
if final_move:
policy = self.build_policy(final_move, False)
history.append(final_move)
policys.append(policy)
state = senv.step(state, final_move)
history.append(state)
self.player.close()
if turns % 2 == 1: # balck turn
value = -value
v = value
if v == 0 or turns <= 10:
if random() > 0.7:
store = True
else:
store = False
else:
store = True
if store:
data = []
for i in range(turns):
k = i * 2
data.append([history[k], policys[i], value])
value = -value
self.save_play_data(idx, data)
self.cur_pipes.append(pipes)
self.remove_play_data()
return v, turns, state, search_tree, store