new_policy.set_state(main_state)
# We need to sync the just copied local weights (from main policy)
# to all the remote workers as well.
trainer.workers.sync_weights()
else:
print("not good enough; will keep learning ...")
# +2 = main + random
result["league_size"] = self.current_opponent + 2
if __name__ == "__main__":
ray.init(num_cpus=args.num_cpus or None, include_dashboard=False)
register_env("open_spiel_env",
lambda _: OpenSpielEnv(pyspiel.load_game(args.env)))
def policy_mapping_fn(agent_id, episode, worker, **kwargs):
# agent_id = [0|1] -> policy depends on episode ID
# This way, we make sure that both policies sometimes play agent0
# (start player) and sometimes agent1 (player to move 2nd).
return "main" if episode.episode_id % 2 == agent_id else "random"
config = {
"env": "open_spiel_env",
"callbacks": SelfPlayCallback,
"model": {
"fcnet_hiddens": [512, 512],
},
"num_sgd_iter": 20,
"num_envs_per_worker": 5,
import pyspiel
import unittest
import ray
import ray.rllib.algorithms.alpha_star as alpha_star
from ray.rllib.env.wrappers.open_spiel import OpenSpielEnv
from ray.rllib.utils.test_utils import (
check_compute_single_action,
check_train_results,
framework_iterator,
)
from ray.tune import register_env
# Connect-4 OpenSpiel env.
register_env("connect_four",
lambda _: OpenSpielEnv(pyspiel.load_game("connect_four")))
class TestAlphaStar(unittest.TestCase):
@classmethod
def setUpClass(cls):
ray.init(num_cpus=20)
@classmethod
def tearDownClass(cls):
ray.shutdown()
def test_alpha_star_compilation(self):
"""Test whether AlphaStar can be built with all frameworks."""
config = (alpha_star.AlphaStarConfig().environment(
env="connect_four").training(
wr = self.win_rates[p] if p in self.win_rates else 0.0
print(f"\t{p}: {wr}")
print("Frozen policies:")
for p in sorted(self.non_trainable_policies):
wr = self.win_rates[p] if p in self.win_rates else 0.0
print(f"\t{p}: {wr}")
print()
if __name__ == "__main__":
ray.init(
num_cpus=args.num_cpus or None,
include_dashboard=False,
)
register_env("open_spiel_env", lambda _: OpenSpielEnv(pyspiel.load_game(args.env)))
def policy_mapping_fn(agent_id, episode, worker, **kwargs):
# At first, only have main play against the random main exploiter.
return "main" if episode.episode_id % 2 == agent_id else "main_exploiter_0"
config = {
"env": "open_spiel_env",
"callbacks": LeagueBasedSelfPlayCallback,
"num_sgd_iter": 20,
"num_envs_per_worker": 5,
"multiagent": {
# Initial policy map: All PPO. This will be expanded
# to more policy snapshots. This is done in the
# custom callback defined above (`LeagueBasedSelfPlayCallback`).
"policies": {
