def do_self_play(board_size, agent1_filename, agent2_filename, num_games,
experience_filename, gpu_frac):
kerasutil.set_gpu_memory_target(gpu_frac)
random.seed(int(time.time()) + os.getpid())
np.random.seed(int(time.time()) + os.getpid())
agent1 = load_agent(agent1_filename)
agent2 = load_agent(agent2_filename)
collector1 = rl.ExperienceCollector()
color1 = Player.black
for i in range(num_games):
print('Simulating game %d/%d...' % (i + 1, num_games))
collector1.begin_episode()
agent1.set_collector(collector1)
if color1 == Player.black:
black_player, white_player = agent1, agent2
else:
white_player, black_player = agent1, agent2
game_record = simulate_game(black_player, white_player, board_size)
if game_record.winner == color1:
print('Agent 1 wins.')
collector1.complete_episode(reward=1)
else:
print('Agent 2 wins.')
collector1.complete_episode(reward=-1)
color1 = color1.other
experience = rl.combine_experience([collector1])
print('Saving experience buffer to %s\n' % experience_filename)
with h5py.File(experience_filename, 'w') as experience_outf:
experience.serialize(experience_outf)
def play_games(args):
agent1_fname, agent2_fname, num_games, board_size, gpu_frac = args
kerasutil.set_gpu_memory_target(gpu_frac)
random.seed(int(time.time()) + os.getpid())
np.random.seed(int(time.time()) + os.getpid())
with h5py.File(agent1_fname, 'r') as agent1f:
agent1 = agent.load_policy_agent(agent1f)
with h5py.File(agent2_fname, 'r') as agent2f:
agent2 = agent.load_policy_agent(agent2f)
wins, losses = 0, 0
color1 = Player.black
for i in range(num_games):
print('Simulating game %d/%d...' % (i + 1, num_games))
if color1 == Player.black:
black_player, white_player = agent1, agent2
else:
white_player, black_player = agent1, agent2
game_record = simulate_game(black_player, white_player, board_size)
if game_record.winner == color1:
print('Agent 1 wins')
wins += 1
else:
print('Agent 2 wins')
losses += 1
print('Agent 1 record: %d/%d' % (wins, wins + losses))
color1 = color1.other
return wins, losses
def play_games(args):
agent1_fname, agent2_fname, num_games, board_size, gpu_frac, temperature = args
kerasutil.set_gpu_memory_target(gpu_frac)
random.seed(int(time.time()) + os.getpid())
np.random.seed(int(time.time()) + os.getpid())
agent1 = load_agent(agent1_fname)
agent1.set_temperature(temperature)
agent1.set_policy('eps-greedy')
agent2 = load_agent(agent2_fname)
agent2.set_temperature(temperature)
agent2.set_policy('eps-greedy')
wins, losses = 0, 0
color1 = Player.black
for i in range(num_games):
print('Simulating game %d/%d...' % (i + 1, num_games))
if color1 == Player.black:
black_player, white_player = agent1, agent2
else:
white_player, black_player = agent1, agent2
game_record = simulate_game(black_player, white_player, board_size)
if game_record.winner == color1:
print('Agent 1 wins')
wins += 1
else:
print('Agent 2 wins')
losses += 1
print('Agent 1 record: %d/%d' % (wins, wins + losses))
color1 = color1.other
return (wins, losses)
def play_games(agent1_fname, agent2_fname,
num_games=480, board_size=19,
gpu_frac=0.95, temperature=0.0):
kerasutil.set_gpu_memory_target(gpu_frac)
agent1 = load_agent(agent1_fname)
agent1.set_temperature(temperature)
agent1.set_policy('eps-greedy')
agent2 = load_agent(agent2_fname)
agent2.set_temperature(temperature)
agent2.set_policy('eps-greedy')
wins, losses = 0, 0
color1 = Player.black
for i in range(num_games):
print('Simulating game %d/%d...' % (i + 1, num_games))
if color1 == Player.black:
black_player, white_player = agent1, agent2
else:
white_player, black_player = agent1, agent2
game_record = simulate_game(black_player, white_player, board_size)
if game_record.winner == color1:
print('Agent 1 wins , time is %s' % (datetime.datetime.now()))
wins += 1
else:
print('Agent 2 wins, time is %s' % (datetime.datetime.now()))
losses += 1
print('Agent 1 record: %d/%d' % (wins, wins + losses))
color1 = color1.other
return wins, losses
def do_self_play(board_size,
agent1_filename,
agent2_filename,
num_games,
temperature,
experience_filename,
chunk=100,
gpu_frac=0.95):
kerasutil.set_gpu_memory_target(gpu_frac)
random.seed(int(time.time()) + os.getpid())
np.random.seed(int(time.time()) + os.getpid())
agent1 = load_agent(agent1_filename)
agent1.set_temperature(temperature)
agent1.set_policy('eps-greedy')
agent2 = load_agent(agent2_filename)
agent2.set_temperature(temperature)
agent2.set_policy('eps-greedy')
color1 = Player.black
times = int(num_games / chunk)
for current_chunk in range(times):
print('Текущая порция %d' % current_chunk)
collector1 = rl.ExperienceCollector()
for i in range(chunk):
print('Симуляция игры %d/%d...' % (i + 1, chunk))
collector1.begin_episode()
agent1.set_collector(collector1)
if color1 == Player.black:
black_player, white_player = agent1, agent2
else:
white_player, black_player = agent1, agent2
game_record = simulate_game(black_player, white_player, board_size)
cnt_moves = len(game_record.moves)
if game_record.winner == color1:
print('Агент 1 выигрывает, время: %s' %
(datetime.datetime.now()))
collector1.complete_episode(reward=1)
else:
print('Агент 2 выигрывает, время: %s' %
(datetime.datetime.now()))
collector1.complete_episode(reward=-1)
print('Количество ходов в игре = ', cnt_moves)
color1 = color1.other
experience = rl.combine_experience([collector1])
print('Saving experience buffer to %s\n' % (experience_filename + str(
(current_chunk + 1) * chunk) + '.h5'))
with h5py.File(
experience_filename + '_' + str(
(current_chunk + 1) * chunk) + '.h5',
'w') as experience_outf:
experience.serialize(experience_outf)
def do_self_play(args):
work_dir, board_size, agent1_num, agent2_num, num_games, experience_filename, gpu_frac, load_args = args
kerasutil.set_gpu_memory_target(gpu_frac)
random.seed(int(time.time()) + os.getpid())
np.random.seed(int(time.time()) + os.getpid())
agent1 = load_agent(agent1_num, work_dir, load_args)
agent2 = load_agent(agent2_num, work_dir, load_args)
collector1 = ExperienceCollector()
collector2 = ExperienceCollector()
ag1b = []
ag2b = []
color1 = Player.black
for i in range(num_games):
print('Simulating game %d/%d...' % (i + 1, num_games))
collector1.begin_episode()
agent1.set_collector(collector1)
collector2.begin_episode()
agent2.set_collector(collector2)
if color1 == Player.black:
black_player, white_player = agent1, agent2
else:
white_player, black_player = agent1, agent2
game_record = simulate_game(black_player, white_player, board_size)
if game_record.winner == color1:
print('Agent 1 wins.')
collector1.complete_episode(reward=game_record.margin)
collector2.complete_episode(reward=-game_record.margin)
elif game_record.winner == color1.other:
print('Agent 2 wins.')
collector1.complete_episode(reward=-game_record.margin)
collector2.complete_episode(reward=game_record.margin)
else:
print('Agents play a draw.')
collector1.complete_episode(reward=0)
collector2.complete_episode(reward=0)
if game_record.winner == Player.black:
black_score = game_record.margin
elif game_record.winner == Player.white:
black_score = -game_record.margin
else:
black_score = 0
if color1 == Player.black:
ag1b.append(black_score)
else:
ag2b.append(black_score)
color1 = color1.other
experience = combine_experience([collector1, collector2])
print('Saving experience buffer to %s\n' % experience_filename)
with h5py.File(experience_filename, 'w') as experience_outf:
experience.serialize(experience_outf)
return (ag1b, ag2b)
def play_games(args):
num_games, board_size, gpu_frac, load_args = args
kerasutil.set_gpu_memory_target(gpu_frac)
random.seed(int(time.time()) + os.getpid())
np.random.seed(int(time.time()) + os.getpid())
agent1 = load_agent(1, load_args.agent1_num, load_args.agent1_dir,
load_args)
agent2 = load_agent(2, load_args.agent2_num, load_args.agent2_dir,
load_args)
ag1b = []
ag2b = []
color1 = Player.black
for i in range(num_games):
print('Simulating game %d/%d...' % (i + 1, num_games))
if color1 == Player.black:
black_player, white_player = agent1, agent2
else:
white_player, black_player = agent1, agent2
game_record = simulate_game(black_player, white_player, board_size)
if game_record.winner == color1:
print('Agent 1 wins.')
elif game_record.winner == color1.other:
print('Agent 2 wins.')
else:
print('Agents play a draw.')
if game_record.winner == Player.black:
black_score = game_record.margin
elif game_record.winner == Player.white:
black_score = -game_record.margin
else:
black_score = 0
if color1 == Player.black:
ag1b.append(black_score)
else:
ag2b.append(black_score)
color1 = color1.other
return (ag1b, ag2b)
