def resume_self_play():
env = TicTacToeEnv()
saves = [f for f in listdir(save_dir) if isfile(join(save_dir, f))]
recent_file = max(saves)
policy = EpsilonGreedy(QConvTicTacToe(env), 0)
opposing_policy = EpsilonGreedy(QConvTicTacToe(env), 1)
self_play = SelfPlay(policy, opposing_policy)
policy.q.policy_net.load_state_dict(torch.load(join(save_dir, recent_file)))
self_play.evaluate_policy(100)
def resume_self_play():
env = Connect4Env()
saves = [f for f in listdir(save_dir) if isfile(join(save_dir, f))]
recent_file = max(saves)
policy = EpsilonGreedy(QLinear(env), 0)
opposing_policy = EpsilonGreedy(QLinear(env), 0) # Acts greedily
self_play = SelfPlay(policy, opposing_policy)
policy.q.policy_net.load_state_dict(torch.load(join(save_dir, recent_file)))
self_play.evaluate_policy(100)
def run_training():
env = TicTacToeEnv()
policy = EpsilonGreedy(QConvTicTacToe(env, buffer_size=5000, batch_size=64), 0.1)
opposing_policy = EpsilonGreedy(
QConvTicTacToe(env), 1
) # Make it not act greedily for the moment- exploration Acts greedily
self_play = SelfPlay(policy, opposing_policy, env=env)
self_play.train_model(20000, resume=False)
print("Training Done")
saved_name = os.path.join(save_dir, datetime.datetime.now().isoformat())
torch.save(self_play.policy.q.policy_net.state_dict(), saved_name)
def run_training():
env = Connect4Env()
# policy = EpsilonGreedy(QConvTicTacToe(env, buffer_size=5000, batch_size=64), 0.1)
policy = MCTreeSearch(
ConvNetConnect4(),
Connect4Env,
temperature_cutoff=3,
iterations=200,
min_memory=64,
)
opposing_policy = EpsilonGreedy(
QConvConnect4(env), 1
) # Make it not act greedily for the moment- exploration Acts greedily
self_play = SelfPlay(policy, opposing_policy, env=env, swap_sides=True)
self_play.train_model(20000, resume=False)
print("Training Done")
saved_name = os.path.join(save_dir, datetime.datetime.now().isoformat())
torch.save(self_play.policy.q.policy_net.state_dict(), saved_name)