def run(self, lr, silent=False):
self.player1 = self.pretrained_player if self.pretrained_player else FCBaselinePlayer(
lr=lr)
if self.opponent is None:
self.player2 = SearchPlayer(
search_depth=3, strategy=OthelloHeuristic.RGRUENER_STRATEGY)
else:
self.player2 = self.opponent
self.simulation = Othello([self.player1, self.player2])
games_per_evaluation = self.games // self.evaluations
start_time = datetime.now()
for episode in range(1, self.evaluations + 1):
# train
self.player1.strategy.train, self.player1.strategy.model.training = True, True # training mode
results, losses = self.simulation.run_simulations(
games_per_evaluation)
self.add_loss(np.mean(losses))
self.add_results((str(self.player2), np.mean(results)))
# evaluate
self.player1.strategy.train, self.player1.strategy.model.training = False, False # eval mode
if self.opponent is None:
score, results, overview = evaluate_against_base_players(
self.player1)
else:
score, results, overview = evaluate_against_base_players(
self.player1, evaluation_players=[self.opponent])
self.add_results(results)
if not silent:
if Printer.print_episode(episode * games_per_evaluation,
self.games,
datetime.now() - start_time):
overview = format_overview(overview)
self.plot_and_save(
"%s vs TRADITIONAL OPPONENT" % (self.player1),
"Train %s vs %s\nGames: %s Evaluations: %s\nTime: %s" %
(self.player1, self.opponent,
episode * games_per_evaluation, self.evaluations,
config.time_diff(start_time)))
self.final_score, self.final_results, self.results_overview = evaluate_against_base_players(
self.player1, silent=False)
return self
def run(self, lr, silent=False):
self.player1 = self.pretrained_player if self.pretrained_player else FCACPlayer(
lr=lr) # ConvACPlayer(lr=lr)
if self.opponent is not None:
self.player2 = self.opponent
self.simulation = Othello([self.player1, self.player2])
games_per_evaluation = self.games // self.evaluations
start_time = datetime.now()
for episode in range(1, self.evaluations + 1):
if self.opponent is None:
self.player2 = choice((RandomPlayer(), NovicePlayer(),
ExperiencedPlayer(deterministic=False)))
self.simulation = Othello([self.player1, self.player2])
# train
self.player1.strategy.train, self.player1.strategy.model.training = True, True # training mode
results, losses = self.simulation.run_simulations(
games_per_evaluation)
self.add_loss(np.mean(losses))
self.add_results(("Training Results", np.mean(results)))
# evaluate
self.player1.strategy.train, self.player1.strategy.model.training = False, False # eval mode
if self.opponent is None:
score, results, overview = evaluate_against_base_players(
self.player1)
else:
score, results, overview = evaluate_against_base_players(
self.player1, evaluation_players=[self.opponent])
self.add_results(results)
if not silent:
if Printer.print_episode(episode * games_per_evaluation,
self.games,
datetime.now() - start_time):
overview = format_overview(overview)
self.plot_and_save(
"%s vs TRADITIONAL OPPONENT" % (self.player1),
"Train %s vs %s\nGames: %s Evaluations: %s\nTime: %s" %
(self.player1, self.opponent,
episode * games_per_evaluation, self.evaluations,
config.time_diff(start_time)))
self.final_score, self.final_results, self.results_overview = evaluate_against_base_players(
self.player1, silent=False)
return self
def run(self, lr, milestones=False, silent=False):
self.player1 = self.pretrained_player if self.pretrained_player else FCBaselinePlayer(lr=lr)
games_per_evaluation = self.games // self.evaluations
start_time = datetime.now()
for episode in range(1, self.evaluations+1):
# If milestones exist, use them with probability p
if self.milestones and random() < 0.2:
self.player2 = choice(self.milestones)
else:
self.player2 = self.player1.copy(shared_weights=True)
# train
self.player1.strategy.train, self.player1.strategy.model.training = True, True # training mode
self.player2.strategy.train = False
self.simulation = Othello([self.player1, self.player2])
results, losses = self.simulation.run_simulations(games_per_evaluation)
self.add_loss(np.mean(losses))
self.add_results(("Self", np.mean(results)))
# evaluate
if episode*games_per_evaluation % 1000 == 0:
self.player1.strategy.train, self.player1.strategy.model.training = False, False # eval mode
score, results, overview = evaluate_against_base_players(self.player1)
self.add_results(results)
if not silent and Printer.print_episode(episode*games_per_evaluation, self.games, datetime.now() - start_time):
self.plot_and_save(
"%s vs SELF" % (self.player1.__str__() + (" milestones" if milestones else "")),
"Train %s vs Self\nGames: %s Evaluations: %s\nTime: %s"
% (self.player1, episode*games_per_evaluation, self.evaluations, config.time_diff(start_time)))
# If x/5th of training is completed, save milestone
if milestones and (self.games / episode * games_per_evaluation) % 5 == 0:
self.milestones.append(self.player1.copy(shared_weights=False))
self.final_score, self.final_results, self.results_overview = evaluate_against_base_players(self.player1, silent=False)
return self
def run(self, lr, silent=False):
self.player1 = self.pretrained_player if self.pretrained_player else FCBaselinePlayer(
lr=lr)
# Player 2 has the same start conditions as Player 1 but does not train
self.player2 = self.player1.copy(shared_weights=False)
self.player2.strategy.train, self.player2.strategy.model.training = False, False # eval mode
games_per_evaluation = self.games // self.evaluations
self.replacements = (0, 0)
start_time = datetime.now()
for episode in range(1, self.evaluations + 1):
# If milestones exist, use them with probability p
if self.milestones and random() < 0.2:
self.player2 = choice(self.milestones)
self.simulation = Othello([self.player1, self.player2])
# train
self.player1.strategy.train, self.player1.strategy.model.training = True, True # training mode
results, losses = self.simulation.run_simulations(
games_per_evaluation)
self.add_loss(np.mean(losses))
self.add_results(("Best", np.mean(results)))
# evaluate
if episode * games_per_evaluation % 1000 == 0:
self.player1.strategy.train, self.player1.strategy.model.training = False, False # eval mode
score, results, overview = evaluate_against_base_players(
self.player1)
self.add_results(results)
if not silent and Printer.print_episode(
episode * games_per_evaluation, self.games,
datetime.now() - start_time):
self.plot_and_save(
"%s vs BEST" % (self.player1.__str__() +
(" milestones" if MILESTONES else "")),
"Train %s vs Best version of self\nGames: %s Evaluations: %s Replacement ratio: %s\nTime: %s"
% (self.player1, episode * games_per_evaluation,
self.evaluations, self.replacements[0] /
self.replacements[1], config.time_diff(start_time)))
if evaluate_against_each_other(self.player1, self.player2,
games=8):
self.player2 = self.player1.copy(shared_weights=False)
self.player2.strategy.train, self.player2.strategy.model.training = False, False
self.replacements = self.replacements[
0] + 1, self.replacements[1] + 1
else:
self.replacements = self.replacements[
0], self.replacements[1] + 1
# If x/5th of training is completed, save milestone
if MILESTONES and (self.games / episode *
games_per_evaluation) % 5 == 0:
self.milestones.append(self.player1.copy(shared_weights=False))
self.milestones[-1].strategy.train = False
self.final_score, self.final_results, self.results_overview = evaluate_against_base_players(
self.player1, silent=False)
return self
