def test_tictactoe_take_win():
gm = GameMaster(get_gdl_for_game("ticTacToe"))
# add two c++ players
gm.add_player(get.get_player("ggtest1"), "xplayer")
gm.add_player(get.get_player("ggtest1"), "oplayer")
str_state = '''
(true (control xplayer))
(true (cell 2 2 x))
(true (cell 3 2 o))
(true (cell 3 3 x))
(true (cell 2 3 o))
(true (cell 3 1 b))
(true (cell 2 1 b))
(true (cell 1 3 b))
(true (cell 1 2 b))
(true (cell 1 1 b)) '''
gm.start(meta_time=30, move_time=2,
initial_basestate=gm.convert_to_base_state(str_state))
# play a single move - should take win
move = gm.play_single_move()
assert str(move[0]) == "(mark 1 1)"
assert str(move[1]) == "noop"
gm.play_to_end(last_move=move)
# check scores/depth make some sense
assert gm.scores['xplayer'] == 100
assert gm.scores['oplayer'] == 0
assert gm.get_game_depth() == 1
def do_game(game, gen, players, meta_time, move_time):
# ensure we have a network
gm = GameMaster(get_gdl_for_game(game), verbose=False)
info = lookup.by_name(game)
for player, role in zip(players, gm.sm.get_roles()):
gm.add_player(player, role)
roles = gm.sm.get_roles()
match_info = create_match_info(meta_time, move_time, len(roles),
game, [p.get_name() for p, r in gm.players], False)
# save match_info
update_match_info(game, match_info)
gm.start(meta_time=meta_time, move_time=move_time)
the_bs = gm.sm.new_base_state()
gm.sm.get_current_state(the_bs)
set_initial_state(match_info, basestate_to_str(info.model, the_bs))
last_move = None
while not gm.finished():
last_move = gm.play_single_move(last_move)
gm.sm.get_current_state(the_bs)
add_to_next(match_info, list(last_move), basestate_to_str(info.model, the_bs))
update_match_info(game, match_info)
gm.play_to_end(last_move)
set_goals(match_info, [gm.get_score(r) for _, r in gm.players])
update_match_info(game, match_info)
def play_game(generation_name, player_clz, num_previous_states=1):
game = "breakthrough"
# ensure we have a network
man = get_manager()
generation_descr = templates.default_generation_desc(
game,
multiple_policy_heads=True,
num_previous_states=num_previous_states)
nn = man.create_new_network(game, "tiny", generation_descr)
man.save_network(nn, generation_name)
nn.summary()
conf = get_default_conf(generation_name,
max_dump_depth=2,
playouts_per_iteration=42)
attrutil.pprint(conf)
gm = GameMaster(lookup.by_name(game))
gm.add_player(player_clz(conf), "white")
gm.add_player(get.get_player("random"), "black")
gm.start(meta_time=30, move_time=15)
last_move = None
while not gm.finished():
last_move = gm.play_single_move(last_move=last_move)
gm.finalise_match(last_move)
return gm
def play(player_white, player_black, move_time=0.5):
gm = GameMaster(lookup.by_name(GAME), verbose=True)
gm.add_player(player_white, "white")
gm.add_player(player_black, "black")
gm.start(meta_time=15, move_time=move_time)
move = None
while not gm.finished():
# print out the board
pretty_board(BOARD_SIZE, gm.sm)
move = gm.play_single_move(last_move=move)
gm.finalise_match(move)
def play_game(config_0, config_1, game_size=11, move_time=10.0):
# add players
gm = GameMaster(get_gdl_for_game("hexLG%s" % game_size))
gm.add_player(PUCTPlayer(config_0), "black")
gm.add_player(PUCTPlayer(config_1), "white")
# play move via gamemaster:
gm.reset()
gm.start(meta_time=15, move_time=move_time)
def remove_gdl(m):
return m.replace("(place ", "").replace(")",
"").strip().replace(' ', '')
move = None
sgf_moves = []
while not gm.finished():
move = gm.play_single_move(last_move=move)
if move[0] == "noop":
ri, str_move = 1, remove_gdl(move[1])
else:
ri, str_move = 0, remove_gdl(move[0])
sgf_moves.append((ri, str_move))
if str_move == "swap":
sgf_moves[0] = (0, swapaxis(sgf_moves[0][1]))
for ri, m in sgf_moves:
print ri, m
x = hashlib.md5(hashlib.sha1("%.5f" %
time.time()).hexdigest()).hexdigest()[:6]
with open("game_%s_%s_%s.sgf" % (config_0.name, config_1.name, x),
"w") as f:
f.write("(;FF[4]EV[null]PB[%s]PW[%s]SZ[%s]GC[game#%s];" %
(config_1.name, config_0.name, game_size, x))
# piece colours are swapped for hexgui from LG
for ri, m in sgf_moves:
f.write("%s[%s];" % ("W" if ri == 0 else "B", m))
f.write(")\n")
def test_not_taking_win():
gm = GameMaster(get_gdl_for_game("breakthrough"))
# add two players
gm.add_player(PUCTPlayer(default_puct_config), "white")
gm.add_player(PUCTPlayer(default_puct_config), "black")
str_state = """ (true (control black))
(true (cellHolds 8 8 black)) (true (cellHolds 8 7 black)) (true (cellHolds 8 2 white))
(true (cellHolds 8 1 white)) (true (cellHolds 7 8 black)) (true (cellHolds 6 7 white))
(true (cellHolds 6 2 white)) (true (cellHolds 6 1 white)) (true (cellHolds 5 4 black))
(true (cellHolds 5 3 black)) (true (cellHolds 5 2 white)) (true (cellHolds 5 1 white))
(true (cellHolds 4 8 black)) (true (cellHolds 4 7 black)) (true (cellHolds 4 1 white))
(true (cellHolds 3 8 black)) (true (cellHolds 3 6 black)) (true (cellHolds 3 2 white))
(true (cellHolds 2 8 black)) (true (cellHolds 2 7 black)) (true (cellHolds 2 2 white))
(true (cellHolds 2 1 white)) (true (cellHolds 1 8 black)) (true (cellHolds 1 7 black))
(true (cellHolds 1 2 white)) (true (cellHolds 1 1 white)) """
gm.start(meta_time=30, move_time=5,
initial_basestate=gm.convert_to_base_state(str_state))
last_move = gm.play_single_move(last_move=None)
assert last_move[1] == "(move 7 8 6 7)"
def play(moves, move_time):
# add players
gm = GameMaster(get_gdl_for_game("breakthroughSmall"))
for role in gm.sm.get_roles():
gm.add_player(PUCTPlayer(conf=compete), role)
sm = lookup.by_name("breakthroughSmall").get_sm()
sm.reset()
pretty_board(sm)
# get some states
joint_move = sm.get_joint_move()
base_state = sm.get_initial_state()
def to_cords(s):
mapping_x_cord = {x0: x1 for x0, x1 in zip('abcdef', '654321')}
return mapping_x_cord[s[0]], s[1]
def f(ri, i):
return sm.legal_to_move(ri, ls.get_legal(i))
lead_role_index = 0
gdl_moves = []
for m in moves:
from_, to_ = map(to_cords, m.split("-"))
gdl_move = []
for ri in range(2):
if ri == lead_role_index:
gdl_role_move = "(move %s %s %s %s)" % (from_[0], from_[1],
to_[0], to_[1])
else:
gdl_role_move = "noop"
ls = sm.get_legal_state(ri)
the_moves = [f(ri, ii) for ii in range(ls.get_count())]
choice = the_moves.index(gdl_role_move)
joint_move.set(ri, ls.get_legal(choice))
gdl_move.append(str(gdl_role_move))
# update state machine
sm.next_state(joint_move, base_state)
sm.update_bases(base_state)
lead_role_index ^= 1
gdl_moves.append(str(gdl_move))
print "%s -> %s" % (m, gdl_move)
pretty_board(sm)
# play move via gamemaster:
gm.reset()
gm.start(meta_time=15,
move_time=move_time,
initial_basestate=sm.get_current_state(),
game_depth=len(moves))
move = gm.play_single_move(last_move=None)
player = gm.get_player(lead_role_index)
move = move[lead_role_index]
move = move.replace("(move", "").replace(")", "")
a, b, c, d = move.split()
mapping_x_cord = {x0: x1 for x0, x1 in zip('654321', 'abcdef')}
next_move = "%s%s-%s%s" % (mapping_x_cord[a], b, mapping_x_cord[c], d)
sm.update_bases(gm.sm.get_current_state())
pretty_board(sm)
print "PLAYED", next_move, player.last_probability
return next_move, player.last_probability
class GameMasterByGame(object):
def __init__(self, game_config):
assert isinstance(game_config, GameConfig)
self.game_config = game_config
game_name = self.game_config.game_name
if game_name == "breakthrough2":
game_name = "breakthrough"
self.gm = GameMaster(get_gdl_for_game(game_name))
# add players
puct_conf = self.get_puct_config()
for role in self.gm.sm.get_roles():
self.gm.add_player(puctplayer.PUCTPlayer(conf=puct_conf), role)
def get_puct_config(self):
multiplier = self.game_config.sims_multiplier
if multiplier == 0:
playouts_per_iteration = 1
else:
playouts_per_iteration = 100 * multiplier
conf = confs.PUCTPlayerConfig(
name="clx",
generation=self.game_config.generation,
verbose=True,
playouts_per_iteration=playouts_per_iteration,
playouts_per_iteration_noop=0,
dirichlet_noise_alpha=-1,
root_expansions_preset_visits=-1,
fpu_prior_discount=0.25,
puct_before_expansions=3,
puct_before_root_expansions=4,
puct_constant_before=3.0,
puct_constant_after=0.75,
choose="choose_temperature",
temperature=1.5,
depth_temperature_max=self.game_config.depth_temperature_max,
depth_temperature_start=4,
depth_temperature_increment=0.25,
depth_temperature_stop=self.game_config.depth_temperature_stop,
random_scale=0.9,
max_dump_depth=2)
return conf
def wtf(self, str_state):
bs = self.gm.convert_to_base_state(str_state)
nn = self.gm.get_player(0).nn
print nn.predict_1(bs.to_list())
def get_move(self, str_state, depth, lead_role_index):
print "GameMasterByGame.get_move", str_state
self.gm.reset()
self.gm.start(
meta_time=120,
move_time=120,
initial_basestate=self.gm.convert_to_base_state(str_state),
game_depth=depth)
move = self.gm.play_single_move(last_move=None)
player = self.gm.get_player(lead_role_index)
return move[
lead_role_index], player.last_probability, self.gm.finished()