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 test_trained():
# simplemcts vs GOOD_GEN
simple = get.get_player("simplemcts")
simple.max_run_time = 0.5
eval_config = confs.PUCTEvaluatorConfig(verbose=True,
puct_constant=0.85,
puct_constant_root=3.0,
dirichlet_noise_pct=-1,
fpu_prior_discount=0.25,
fpu_prior_discount_root=0.15,
choose="choose_temperature",
temperature=2.0,
depth_temperature_max=10.0,
depth_temperature_start=0,
depth_temperature_increment=0.75,
depth_temperature_stop=1,
random_scale=1.0,
batch_size=1,
max_dump_depth=1)
puct_config = confs.PUCTPlayerConfig("gzero", True, 200, 0, GOOD_GEN1,
eval_config)
attrutil.pprint(puct_config)
puct_player = PUCTPlayer(puct_config)
play(simple, puct_player)
def test_random():
# add two players
# simplemcts vs RANDOM_GEN
pymcs = get.get_player("simplemcts")
pymcs.max_run_time = 0.25
eval_config = templates.base_puct_config(verbose=True, max_dump_depth=1)
puct_config = confs.PUCTPlayerConfig("gzero", True, 100, 0, RANDOM_GEN,
eval_config)
attrutil.pprint(puct_config)
puct_player = PUCTPlayer(puct_config)
play(pymcs, puct_player)
def get_puct_config(gen, **kwds):
eval_config = confs.PUCTEvaluatorConfig(verbose=True,
puct_constant=0.85,
puct_constant_root=3.0,
dirichlet_noise_pct=-1,
fpu_prior_discount=0.25,
fpu_prior_discount_root=0.15,
choose="choose_temperature",
temperature=2.0,
depth_temperature_max=10.0,
depth_temperature_start=0,
depth_temperature_increment=0.75,
depth_temperature_stop=1,
random_scale=1.0,
max_dump_depth=2,
top_visits_best_guess_converge_ratio=0.8,
think_time=2.0,
converged_visits=2000,
batch_size=32)
config = confs.PUCTPlayerConfig(name="puct",
verbose=True,
generation=gen,
playouts_per_iteration=-1,
playouts_per_iteration_noop=0,
evaluator_config=eval_config)
for k, v in kwds.items():
updated = False
if at.has(eval_config, k):
updated = True
setattr(eval_config, k, v)
if at.has(config, k):
updated = True
setattr(config, k, v)
if not updated:
log.warning("Unused setting %s:%s" % (k, v))
return config
def config(gen):
return confs.PUCTPlayerConfig(name=gen,
generation=gen,
verbose=True,
playouts_per_iteration=-1,
playouts_per_iteration_noop=0,
dirichlet_noise_alpha=-1,
root_expansions_preset_visits=1,
puct_before_expansions=3,
puct_before_root_expansions=5,
puct_constant_before=3.0,
puct_constant_after=0.75,
choose="choose_temperature",
temperature=1.5,
depth_temperature_max=3.0,
depth_temperature_start=0,
depth_temperature_increment=0.5,
depth_temperature_stop=6,
random_scale=1.00,
fpu_prior_discount=-1,
max_dump_depth=2)
def test_fast_plays():
''' very fast rollouts, basically this config of puct player is a policy player '''
gm = GameMaster(get_gdl_for_game("breakthrough"))
import attr
conf = confs.PUCTPlayerConfig(**attr.asdict(default_puct_config))
conf.verbose = False
# just checking that we haven't modified default
assert not conf.verbose and default_puct_config.verbose
conf.playouts_per_iteration = 1
conf.playouts_per_iteration_noop = 0
conf.dirichlet_noise_alpha = -1
print conf
# add two players
white = PUCTPlayer(conf=conf)
black = PUCTPlayer(conf=conf)
gm.add_player(white, "white")
gm.add_player(black, "black")
acc_black_score = 0
acc_red_score = 0
s = time.time()
for _ in range(ITERATIONS):
gm.start(meta_time=30, move_time=15)
gm.play_to_end()
acc_black_score += gm.scores["black"]
acc_red_score += gm.scores["white"]
print gm.get_game_depth()
print "time taken", time.time() - s
print "white_score", gm.players_map["white"].name, acc_red_score
print "black_score", gm.players_map["black"].name, acc_black_score
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
tf.logging.set_verbosity(tf.logging.ERROR)
compete = confs.PUCTPlayerConfig(
name="bt_config",
##generation="x1_109",
generation="x1_132",
verbose=True,
playouts_per_iteration=-1,
playouts_per_iteration_noop=0,
dirichlet_noise_alpha=-1,
root_expansions_preset_visits=-1,
puct_before_expansions=3,
puct_before_root_expansions=5,
puct_constant_before=3.0,
puct_constant_after=0.75,
choose="choose_top_visits",
temperature=1.5,
depth_temperature_max=3.0,
depth_temperature_start=0,
depth_temperature_increment=0.5,
depth_temperature_stop=4,
random_scale=1.00,
max_dump_depth=6)
def pretty_board(sm):
' pretty print board current state of match '
from ggplib.util.symbols import SymbolFactory
def puct_config_template(generation, name="default"):
configs = dict(
default=confs.PUCTPlayerConfig(name="default",
verbose=True,
playouts_per_iteration=2400,
playouts_per_iteration_noop=800,
dirichlet_noise_alpha=0.05,
puct_before_expansions=3,
puct_before_root_expansions=5,
puct_constant_before=3.0,
puct_constant_after=0.75,
choose="choose_top_visits",
max_dump_depth=2),
test=confs.PUCTPlayerConfig(name="test",
verbose=True,
playouts_per_iteration=42,
playouts_per_iteration_noop=0,
dirichlet_noise_alpha=0.03,
puct_before_expansions=3,
puct_before_root_expansions=5,
puct_constant_before=3.0,
puct_constant_after=0.75,
choose="choose_top_visits",
max_dump_depth=2),
compete=confs.PUCTPlayerConfig(name="compete",
verbose=True,
playouts_per_iteration=400,
playouts_per_iteration_noop=400,
dirichlet_noise_alpha=-1,
puct_before_expansions=3,
puct_before_root_expansions=5,
puct_constant_before=3.0,
puct_constant_after=1.00,
temperature=2.0,
depth_temperature_max=10.0,
depth_temperature_start=4,
depth_temperature_increment=0.25,
depth_temperature_stop=100,
random_scale=0.75,
choose="choose_temperature",
max_dump_depth=2),
compete2=confs.PUCTPlayerConfig(name="compete2",
verbose=True,
playouts_per_iteration=800,
playouts_per_iteration_noop=800,
root_expansions_preset_visits=7,
resign_score_value=0.05,
playouts_per_iteration_resign=25,
dirichlet_noise_alpha=-1,
puct_before_expansions=3,
puct_before_root_expansions=5,
puct_constant_before=3.0,
puct_constant_after=1.00,
temperature=1.0,
depth_temperature_max=2.5,
depth_temperature_start=8,
depth_temperature_increment=0.1,
depth_temperature_stop=60,
random_scale=0.65,
choose="choose_temperature",
max_dump_depth=2),
policy=confs.PUCTPlayerConfig(name="policy-test",
verbose=True,
playouts_per_iteration=0,
playouts_per_iteration_noop=0,
dirichlet_noise_alpha=-1,
choose="choose_top_visits",
max_dump_depth=1),
policy_compete=confs.PUCTPlayerConfig(name="policy_compete",
verbose=True,
playouts_per_iteration=0,
playouts_per_iteration_noop=0,
dirichlet_noise_alpha=-1,
temperature=1.0,
depth_temperature_max=2.5,
depth_temperature_start=8,
depth_temperature_increment=0.1,
depth_temperature_stop=60,
random_scale=0.65,
choose="choose_temperature",
max_dump_depth=1),
max_score=confs.PUCTPlayerConfig(name="max-score",
verbose=True,
playouts_per_iteration=1,
playouts_per_iteration_noop=0,
dirichlet_noise_alpha=-1,
puct_constant_before=0,
puct_constant_after=0,
choose="choose_top_visits",
max_dump_depth=2),
compare=confs.PUCTPlayerConfig(name="compare",
verbose=True,
playouts_per_iteration=150,
playouts_per_iteration_noop=1,
dirichlet_noise_alpha=0.03,
puct_before_expansions=3,
puct_before_root_expansions=5,
puct_constant_before=3.0,
puct_constant_after=0.75,
choose="choose_top_visits",
max_dump_depth=2))
conf = configs[name]
conf.generation = generation
return conf
from ggplib.db.helper import get_gdl_for_game
from ggpzero.defs import confs
from ggpzero.player.puctplayer import PUCTPlayer
import py.test
ITERATIONS = 1
current_gen = "x6_111"
# first in the run, completely random weights
random_gen = "x5_0"
default_puct_config = confs.PUCTPlayerConfig(generation=current_gen,
playouts_per_iteration=42,
playouts_per_iteration_noop=1)
def setup():
import tensorflow as tf
from ggplib.util.init import setup_once
setup_once()
from ggpzero.util.keras import init
init()
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
tf.logging.set_verbosity(tf.logging.ERROR)