def run_games(config):
game = Othello()
model = ""
x = config.iterations
while(x != 0):
x -= 1
models = sorted(glob.glob(config.data.model_location+"*.h5"))
if model == "":
model = models[-1]
print("Loading new model: %s" % util.getPlayerName(model))
ai = AIPlayer(config.buffer_size, config.game.simulation_num_per_move, model=model)
elif models[-1] != model:
model = models[-1]
print("Loading new model: %s" % util.getPlayerName(model))
ai.load(model)
start=time()
for j in range(config.nb_game_in_file):
util.print_progress_bar(j, config.nb_game_in_file, start=start)
side = -1
turn = 1
while not game.game_over():
ai.tau = config.game.tau_1
if config.game.tau_swap < turn:
ai.tau = config.game.tau_2
t = ai.pick_move(game, side)
game.play_move(t[0], t[1], side)
side *= -1
turn += 1
ai.update_buffer(game.get_winner())
game.reset_board()
#print("Average Game Time: ", (time()-start)/(config.nb_game_in_file))
util.print_progress_bar(config.nb_game_in_file, config.nb_game_in_file, start=start)
save_games(config, ai.buffer)
t.join()
def train(ai, config):
loaded_files = []
x = config.iterations
i = len(glob.glob(config.data.model_location + "*.h5"))
while (x != 0):
if i > config.iter3:
ai.update_lr(config.learning_rate3)
elif i > config.iter2:
ai.update_lr(config.learning_rate2)
else:
ai.update_lr(config.learning_rate1)
loaded_files = load_games(ai, loaded_files, config)
length = 0
start = time()
print("Iteration %04d" % i)
progress_start = (
i - 2
) * config.min_new_game_files + config.min_game_files if i > 1 else 0
progress_end = (i -
1) * config.min_new_game_files + config.min_game_files
util.print_progress_bar(0, progress_end - progress_start, start=start)
while (len(loaded_files) < config.min_game_files
or ((len(loaded_files) - config.min_game_files) //
config.min_new_game_files) + 1 < i):
if length != len(loaded_files):
length = len(loaded_files)
util.print_progress_bar(length - progress_start,
progress_end - progress_start,
start=start)
sleep(5)
loaded_files = load_games(ai, loaded_files, config)
util.print_progress_bar(progress_end - progress_start,
progress_end - progress_start,
start=start)
print("Training for %d batches on %d samples" %
(config.batches_per_iter, len(ai.buffer.buffer)))
start = time()
history = ai.train_batches(config.batch_size, config.batches_per_iter,
config.verbose)
for val in history.history.keys():
print("%s: %0.4f" % (val, history.history[val][-1]))
if i % config.save_model_cycles == 0:
ai.save(config.data.model_location + "model_" + str(i) + ".h5")
file = open(
config.data.history_location + "hist_" + str(i) + ".pickle", 'wb')
pickle.dump(pickle.dumps(history.history), file)
file.close()
print("Iteration Time: %0.2f" % (time() - start))
x -= 1
i += 1
def train(ai, config):
loaded_files = []
x = config.iterations
i = len(glob.glob(config.data.model_location + "*.h5"))
loaded_files, _ = load_games(ai, loaded_files, config)
while (x != 0):
if i > config.iter3:
ai.update_lr(config.learning_rate3)
elif i > config.iter2:
ai.update_lr(config.learning_rate2)
else:
ai.update_lr(config.learning_rate1)
loaded_files, diff = load_games(ai, loaded_files, config)
total_diff = diff
start = time()
print("Iteration %04d" % i)
end = config.min_new_game_files if i > 0 else config.min_game_file
util.print_progress_bar(0, end, start=start)
while (total_diff < end):
if diff > 0:
total_diff += diff
util.print_progress_bar(total_diff, end, start=start)
sleep(5)
loaded_files, diff = load_games(ai, loaded_files, config)
util.print_progress_bar(end, end, start=start)
print("Training for %d batches on %d samples" %
(config.batches_per_iter, len(ai.buffer.buffer)))
start = time()
history = ai.train_batches(config.batch_size, config.batches_per_iter,
config.verbose)
for val in history.history.keys():
print("%s: %0.4f" % (val, history.history[val][-1]))
if i % config.save_model_cycles == 0:
ai.save("%smodel_%04d.h5" % (config.data.model_location, i))
file = open("%shist_%04d.pickle" % (config.data.history_location, i),
'wb')
pickle.dump(pickle.dumps(history.history), file)
file.close()
print("Iteration Time: %0.2f" % (time() - start))
x -= 1
i += 1
def train(ai, config):
loaded_files = []
x = config.iterations
i = len(glob.glob(config.data.model_location+"*.h5"))
loaded_files, _ = load_games(ai, loaded_files, config)
while(x != 0):
if i > config.iter3:
ai.update_lr(config.learning_rate3)
elif i > config.iter2:
ai.update_lr(config.learning_rate2)
else:
ai.update_lr(config.learning_rate1)
loaded_files, diff = load_games(ai, loaded_files, config)
total_diff = diff
start = time()
print("Iteration %04d"%i)
end = config.min_new_game_files if i> 0 else config.min_game_file
util.print_progress_bar(0, end, start=start)
while(total_diff < end):
if diff > 0:
total_diff += diff
util.print_progress_bar(total_diff, end, start=start)
sleep(5)
loaded_files, diff = load_games(ai, loaded_files, config)
util.print_progress_bar(end, end, start=start)
print("Training for %d batches on %d samples" % (config.batches_per_iter, len(ai.buffer.buffer)))
start = time()
history = ai.train_batches(config.batch_size, config.batches_per_iter, config.verbose)
for val in history.history.keys():
print("%s: %0.4f" % (val, history.history[val][-1]))
if i % config.save_model_cycles == 0:
ai.save("%smodel_%04d.h5" % (config.data.model_location, i))
file = open("%shist_%04d.pickle" % (config.data.history_location, i), 'wb')
pickle.dump(pickle.dumps(history.history), file)
file.close()
print("Iteration Time: %0.2f" % (time()-start))
x -= 1
i += 1
def run_games(config):
game = Othello()
model_1 = ""
model_2 = ""
p1, new_1 = create_player(config.model_1, model_1, config)
p2, new_2 = create_player(config.model_2, model_2, config)
i = len(glob.glob(config.data.model_location+"*.h5"))
avg_wins = []
while True:
i += 1
new_1 = load_player(p1, config.model_1, model_1, config)
new_2 = load_player(p2, config.model_2, model_2, config)
while((config.model_1 == "newest" and new_1 == model_1) or (config.model_2 == "newest" and new_2 == model_2)):
#print("Waiting on new model. Sleeping for 1 minute.")
sleep(60)
new_1 = load_player(p1, config.model_1, model_1, config)
new_2 = load_player(p2, config.model_2, model_2, config)
model_1 = new_1
model_2 = new_2
wins = 0
losses = 0
ties = 0
print("Iteration %04d"%i)
print("Playing %d games with %d simulations per move" % (config.game_num, config.game.simulation_num_per_move))
start=time()
for j in range(config.game_num):
util.print_progress_bar(j, config.game_num, start=start)
side = -1
turn = 1
while not game.game_over():
tau = config.game.tau_1
if config.game.tau_swap < turn:
tau = config.game.tau_2
if config.model_1 != "random":
p1.tau =tau
if config.model_2 != "random":
p2.tau = tau
if j % 2 == 0:
if side == -1:
t = p1.pick_move(game, side)
else:
t = p2.pick_move(game, side)
else:
if side == 1:
t = p1.pick_move(game, side)
else:
t = p2.pick_move(game, side)
game.play_move(t[0], t[1], side)
side *= -1
turn += 1
if game.get_winner() == 0:
ties += 1
elif j % 2 == 0 and game.get_winner() == -1:
wins += 1
elif j % 2 == 1 and game.get_winner() == 1:
wins += 1
else:
losses += 1
game.reset_board()
util.print_progress_bar(config.game_num, config.game_num, start=start)
print("%s vs %s: (%0.2f%% wins|%0.2f%% ties|%0.2f%% losses) of %d games" % (config.model_1, config.model_2,
100*wins/config.game_num, 100*ties/config.game_num, 100*losses/config.game_num, config.game_num))
avg_wins.append(100*wins/config.game_num)
if len(avg_wins) > config.rolling_avg_amount:
avg_wins = avg_wins[-1*config.rolling_avg_amount:]
print("Average Win Percent: %0.2f%%" % (sum(avg_wins)/float(len(avg_wins))))
if not (config.repeat_with_new_model and (config.model_1 == "newest" or config.model_2 == "newest")):
break
def run_games(config):
game = Othello()
model_1 = ""
model_2 = ""
p1, new_1 = create_player(config.model_1, model_1, config)
p2, new_2 = create_player(config.model_2, model_2, config)
if config.model_1 == "newest" or config.model_2 == "newest":
i = len(glob.glob(config.data.model_location+"*.h5"))-1
else:
i = 0
avg_wins = []
while True:
i += 1
new_1 = load_player(p1, config.model_1, model_1, config)
new_2 = load_player(p2, config.model_2, model_2, config)
while((config.model_1 == "newest" and new_1 == model_1) or (config.model_2 == "newest" and new_2 == model_2)):
#print("Waiting on new model. Sleeping for 1 minute.")
sleep(60)
new_1 = load_player(p1, config.model_1, model_1, config)
new_2 = load_player(p2, config.model_2, model_2, config)
model_1 = new_1
model_2 = new_2
wins = 0
losses = 0
ties = 0
print("Iteration %04d"%i)
print("Playing games between %s and %s" % (config.model_1, config.model_2))
print("Playing %d games with %d simulations per move" % (config.game_num, config.game.simulation_num_per_move))
start=time()
for j in range(config.game_num):
util.print_progress_bar(j, config.game_num, start=start)
side = -1
turn = 1
while not game.game_over():
tau = config.game.tau_1
if config.game.tau_swap < turn:
tau = config.game.tau_2
if config.model_1 != "random":
p1.tau =tau
if config.model_2 != "random":
p2.tau = tau
if j % 2 == 0:
if side == -1:
t = p1.pick_move(game, side)
else:
t = p2.pick_move(game, side)
else:
if side == 1:
t = p1.pick_move(game, side)
else:
t = p2.pick_move(game, side)
game.play_move(t[0], t[1], side)
side *= -1
turn += 1
if game.get_winner() == 0:
ties += 1
savePerformance(config, model_1, model_2, 0, 1, 0)
elif (j % 2 == 0 and game.get_winner() == -1) or (j % 2 == 1 and game.get_winner() == 1):
wins += 1
savePerformance(config, model_1, model_2, 1, 0, 0)
else:
losses += 1
savePerformance(config, model_1, model_2, 0, 0, 1)
game.reset_board()
util.print_progress_bar(config.game_num, config.game_num, start=start)
print("%s vs %s: (%0.2f%% wins|%0.2f%% ties|%0.2f%% losses) of %d games" % (config.model_1, config.model_2,
100*wins/config.game_num, 100*ties/config.game_num, 100*losses/config.game_num, config.game_num))
avg_wins.append(100*wins/config.game_num)
if len(avg_wins) > config.rolling_avg_amount:
avg_wins = avg_wins[-1*config.rolling_avg_amount:]
print("Average Win Percent: %0.2f%%" % (sum(avg_wins)/float(len(avg_wins))))
if not (config.repeat_with_new_model and (config.model_1 == "newest" or config.model_2 == "newest")):
break
def calc_ranking(config):
models = sorted(glob.glob(config.data.model_location + "*.h5"))
players = []
for i, model in enumerate(models):
if i % config.model_skip == 0 or i == len(models):
players.append(model)
wtl = np.zeros((len(players), 3))
win_matrix = np.zeros((len(players), len(players)))
game = Othello()
king_index = len(players) - 1
king = AIPlayer(0,
config.game.simulation_num_per_move,
train=False,
model=players[king_index],
tau=config.game.tau_1)
challenger = AIPlayer(0,
config.game.simulation_num_per_move,
train=False,
model=players[0],
tau=config.game.tau_1)
total_games = (config.game_num_per_model * (len(players))) // 2
played_games = 0
start = time()
print("Playing king of the hill with %d players and %d games per player" %
(len(players), config.game_num_per_model))
if config.game_num_per_model < len(players):
print(
"We suggest that you increase games per player to be greater than players"
)
for i in range(math.ceil(total_games / (len(players) - 1))):
AIPlayer.clear()
king_index = getKingIndex(win_matrix)
if king_index == -1:
king_index = (len(players) - 1) - i % len(players)
msg = "No King Yet"
else:
msg = "King is " + os.path.basename(
players[king_index]).split(".")[0]
king.load(players[king_index])
if config.print_king:
print(msg.ljust(90))
for j in range(len(players)):
util.print_progress_bar(played_games, total_games, start=start)
if j == king_index:
continue
challenger.load(players[j])
if random.random() < 0.5:
king_side = -1
p1 = king
p2 = challenger
else:
king_side = 1
p1 = challenger
p2 = king
side = -1
turn = 1
while not game.game_over():
tau = config.game.tau_1
if config.game.tau_swap < turn:
tau = config.game.tau_2
p1.tau = tau
p2.tau = tau
if side == -1:
t = p1.pick_move(game, side)
else:
t = p2.pick_move(game, side)
game.play_move(t[0], t[1], side)
side *= -1
turn += 1
if game.get_winner() == king_side:
win_matrix[king_index, j] += 1
wtl[king_index, 0] += 1
wtl[j, 2] += 1
elif game.get_winner() == -1 * king_side:
win_matrix[j, king_index] += 1
wtl[king_index, 2] += 1
wtl[j, 0] += 1
else:
win_matrix[king_index, j] += 0.5
win_matrix[j, king_index] += 0.5
wtl[king_index, 1] += 1
wtl[j, 1] += 1
game.reset_board()
played_games += 1
if played_games == total_games:
break
util.print_progress_bar(total_games, total_games, start=start)
try:
params = choix.ilsr_pairwise_dense(win_matrix)
print("\nRankings:")
for i, player in enumerate(np.argsort(params)[::-1]):
print(
"%d. %s (expected %d) with %0.2f rating and results of %d-%d-%d"
% (i + 1, os.path.basename(players[player]).split(".")[0],
len(players) - player, params[player], wtl[player, 0],
wtl[player, 1], wtl[player, 2]))
print(
"\n(Rating Diff, Winrate) -> (0.5, 62%), (1, 73%), (2, 88%), (3, 95%), (5, 99%)"
)
except Exception:
print("\n Not Enough data to calculate rankings")
print("\nWin Matrix:")
print(win_matrix)
print("\nResults:")
for player in range(win_matrix.shape[0]):
print("%s results of %d-%d-%d" %
(os.path.basename(players[player]).split(".")[0],
wtl[player, 0], wtl[player, 1], wtl[player, 2]))
def calc_ranking(config):
models = sorted(glob.glob(config.data.model_location + "*.h5"))
players = []
for i, model in enumerate(models):
if i % config.model_skip == 0 or i == len(models):
players.append(model)
wtl = np.zeros((len(players), 3))
win_matrix = np.zeros((len(players), len(players)))
game = Othello()
##give every player a random order to play games against opponents
order = []
for i in range(len(players)):
nums = [x for x in range(len(players))]
nums.remove(i)
random.shuffle(nums)
order.append(nums)
p1 = AIPlayer(1, config.game.simulation_num_per_move, model=players[0])
p2 = AIPlayer(1,
config.game.simulation_num_per_move,
model=players[order[0][0]])
start = time()
print(
"Playing random round robin with %d players and %d games per player" %
(len(players), config.game_num_per_model))
for i in range(config.game_num_per_model // 2):
util.print_progress_bar(i, config.game_num_per_model // 2, start=start)
ordering = [x for x in range(len(players))]
random.shuffle(ordering)
for j in ordering:
AIPlayer.clear()
x = i
if x >= len(order[j]):
x %= len(order[j])
if x == 0:
random.shuffle(order[j])
p1.load(players[j])
p2.load(players[order[j][x]])
side = -1
turn = 1
while not game.game_over():
tau = config.game.tau_1
if config.game.tau_swap < turn:
tau = config.game.tau_2
p1.tau = tau
p2.tau = tau
if side == -1:
t = p1.pick_move(game, side)
else:
t = p2.pick_move(game, side)
game.play_move(t[0], t[1], side)
side *= -1
turn += 1
if game.get_winner() == -1:
win_matrix[j, order[j][x]] += 1
wtl[j, 0] += 1
wtl[order[j][x], 2] += 1
elif game.get_winner() == 1:
win_matrix[order[j][x], j] += 1
wtl[j, 2] += 1
wtl[order[j][x], 0] += 1
else:
win_matrix[j, order[j][x]] += 0.5
win_matrix[order[j][x], j] += 0.5
wtl[j, 1] += 1
wtl[order[j][x], 1] += 1
game.reset_board()
util.print_progress_bar(config.game_num_per_model // 2,
config.game_num_per_model // 2,
start=start)
params = choix.ilsr_pairwise_dense(win_matrix)
print("\nRankings:")
for i, player in enumerate(np.argsort(params)[::-1]):
print(
"%d. %s (expected %d) with %0.2f rating and results of %d-%d-%d" %
(i + 1, os.path.basename(players[player]), len(players) - player,
params[player], wtl[player, 0], wtl[player, 1], wtl[player, 2]))
print(
"\n(Rating Diff, Winrate) -> (0.5, 62%), (1, 73%), (2, 88%), (3, 95%), (5, 99%)"
)
def calc_ranking(config):
models = sorted(glob.glob(config.data.model_location+"*.h5"))
players = []
for i, model in enumerate(models):
if i % config.model_skip == 0 or i == len(models):
players.append(model)
wtl = np.zeros((len(players), len(players), 3))
win_matrix = np.zeros((len(players),len(players)))
game = Othello()
challenger1 = AIPlayer(0, config.game.simulation_num_per_move, train=False, model=players[-1], tau=config.game.tau_1)
challenger2 = AIPlayer(0, config.game.simulation_num_per_move, train=False, model=players[0], tau=config.game.tau_1)
total_games = (config.game_num_per_model * (len(players)))//2
played_games = 0
finished = False
start = time()
print("Ranking with %d players and %d games per player" % (len(players), config.game_num_per_model))
if config.game_num_per_model < len(players):
print("We suggest that you increase games per player to be greater than players")
for i in itertools.count():
ranks = getRankings(win_matrix)
if len(ranks) == 0:
msg = "No Clear Best Yet"
else:
msg = "Current Best is "+util.getPlayerName(players[ranks[-1]])
if config.print_best:
print(msg.ljust(90))
for j in range(len(players)):
util.print_progress_bar(played_games, total_games, start=start)
challenger1_index = getLeastPlayed(win_matrix, j)
AIPlayer.clear()
challenger1.load(players[challenger1_index])
challenger2.load(players[j])
if random.random() < 0.5:
challenger1_side = -1
p1 = challenger1
p2 = challenger2
else:
challenger1_side = 1
p1 = challenger2
p2 = challenger1
side = -1
turn = 1
while not game.game_over():
tau = config.game.tau_1
if config.game.tau_swap < turn:
tau = config.game.tau_2
p1.tau = tau
p2.tau = tau
if side == -1:
t = p1.pick_move(game, side)
else:
t = p2.pick_move(game, side)
game.play_move(t[0], t[1], side)
side *= -1
turn += 1
if game.get_winner() == challenger1_side:
win_matrix[challenger1_index,j] += 1
wtl[challenger1_index, j,0] += 1
elif game.get_winner() == -1*challenger1_side:
win_matrix[j, challenger1_index] += 1
wtl[challenger1_index, j,2] += 1
else:
win_matrix[challenger1_index,j] += 0.5
win_matrix[j, challenger1_index] += 0.5
wtl[challenger1_index, j, 1] += 1
game.reset_board()
played_games += 1
if played_games >= total_games:
finished = True
break
saveWTL(config, players, wtl)
wtl = np.zeros((len(players), len(players), 3))
if finished:
break
util.print_progress_bar(total_games, total_games, start=start)
print("\n",[util.getPlayerName(player) for player in players])
print("\nWin Matrix(row beat column):")
print(win_matrix)
try:
with np.errstate(divide='ignore', invalid='ignore'):
params = choix.ilsr_pairwise_dense(win_matrix)
print("\nRankings:")
for i, player in enumerate(np.argsort(params)[::-1]):
print("%d. %s (expected %d) with %0.2f rating"%
(i+1, util.getPlayerName(players[player]), len(players)-player, params[player]))
print("\n(Rating Diff, Winrate) -> (0.5, 62%), (1, 73%), (2, 88%), (3, 95%), (5, 99%)")
except Exception:
print("\nNot Enough data to calculate rankings")
def calc_ranking(config):
models = sorted(glob.glob(config.data.model_location + "*.h5"))
players = []
for i, model in enumerate(models):
if i % config.model_skip == 0 or i == len(models):
players.append(model)
wtl = np.zeros((len(players), len(players), 3))
win_matrix = np.zeros((len(players), len(players)))
game = Othello()
challenger1 = AIPlayer(0,
config.game.simulation_num_per_move,
train=False,
model=players[-1],
tau=config.game.tau_1)
challenger2 = AIPlayer(0,
config.game.simulation_num_per_move,
train=False,
model=players[0],
tau=config.game.tau_1)
total_games = (config.game_num_per_model * (len(players))) // 2
played_games = 0
finished = False
start = time()
print("Ranking with %d players and %d games per player" %
(len(players), config.game_num_per_model))
if config.game_num_per_model < len(players):
print(
"We suggest that you increase games per player to be greater than players"
)
for i in itertools.count():
ranks = getRankings(win_matrix)
if len(ranks) == 0:
msg = "No Clear Best Yet"
else:
msg = "Current Best is " + util.getPlayerName(players[ranks[-1]])
if config.print_best:
print(msg.ljust(90))
for j in range(len(players)):
util.print_progress_bar(played_games, total_games, start=start)
challenger1_index = getLeastPlayed(win_matrix, j)
AIPlayer.clear()
challenger1.load(players[challenger1_index])
challenger2.load(players[j])
if random.random() < 0.5:
challenger1_side = -1
p1 = challenger1
p2 = challenger2
else:
challenger1_side = 1
p1 = challenger2
p2 = challenger1
side = -1
turn = 1
while not game.game_over():
tau = config.game.tau_1
if config.game.tau_swap < turn:
tau = config.game.tau_2
p1.tau = tau
p2.tau = tau
if side == -1:
t = p1.pick_move(game, side)
else:
t = p2.pick_move(game, side)
game.play_move(t[0], t[1], side)
side *= -1
turn += 1
if game.get_winner() == challenger1_side:
win_matrix[challenger1_index, j] += 1
wtl[challenger1_index, j, 0] += 1
elif game.get_winner() == -1 * challenger1_side:
win_matrix[j, challenger1_index] += 1
wtl[challenger1_index, j, 2] += 1
else:
win_matrix[challenger1_index, j] += 0.5
win_matrix[j, challenger1_index] += 0.5
wtl[challenger1_index, j, 1] += 1
game.reset_board()
played_games += 1
if played_games >= total_games:
finished = True
break
saveWTL(config, players, wtl)
wtl = np.zeros((len(players), len(players), 3))
if finished:
break
util.print_progress_bar(total_games, total_games, start=start)
print("\n", [util.getPlayerName(player) for player in players])
print("\nWin Matrix(row beat column):")
print(win_matrix)
try:
with np.errstate(divide='ignore', invalid='ignore'):
params = choix.ilsr_pairwise_dense(win_matrix)
print("\nRankings:")
for i, player in enumerate(np.argsort(params)[::-1]):
print("%d. %s (expected %d) with %0.2f rating" %
(i + 1, util.getPlayerName(
players[player]), len(players) - player, params[player]))
print(
"\n(Rating Diff, Winrate) -> (0.5, 62%), (1, 73%), (2, 88%), (3, 95%), (5, 99%)"
)
except Exception:
print("\nNot Enough data to calculate rankings")
