class AIMatchmaker(gym.Env):
metadata = {'render.modes': None}
def __init__(self,
all_stats,
all_opps,
all_elos,
game_path,
model_dir,
base_port=50000,
my_port=50001,
image_based=False,
level_path=None,
env_p=3,
starting_elo=None,
K=16,
D=5.,
time_reward=-0.003,
matchmaking_mode=0,
elo_log_interval=10000,
win_loss_ratio=[0, 0]):
super(AIMatchmaker, self).__init__()
self.all_stats = combine_winrates(all_stats)
self.all_opps = all_opps
self.all_elos = all_elos
self.model_dir = model_dir
self.agent_elo = starting_elo if starting_elo != None else self.all_elos[
0]
self.env = TankEnv(game_path,
opp_fp_and_elo=[],
game_port=base_port,
my_port=my_port,
image_based=image_based,
level_path=level_path,
p=env_p,
time_reward=time_reward)
self.action_space = self.env.action_space
self.observation_space = self.env.observation_space
self.K = K
self.D = D
self.my_port = my_port
self.mm = matchmaking_mode
self.uncounted_games = np.array([0, 0], dtype=np.uint32)
self.counted_game_sets = 0
self.win_loss_ratio = np.array(win_loss_ratio, dtype=np.uint32)
self.started = False
self.next_opp()
self.elo_log_interval = elo_log_interval
self.num_steps = 0
self.elo_log = []
def next_opp(self):
weights = np.zeros((len(self.all_elos)), dtype=np.float32)
if self.mm == 1:
# ELO based matchmaking, where ELOs closer to agent ELo is prefered (but not guarenteed)
weights += np.array([
weight_func(elo - self.agent_elo, self.D)
for elo in self.all_elos
],
dtype=np.float32)
if any(self.win_loss_ratio):
while all(self.uncounted_games >= self.win_loss_ratio):
self.uncounted_games -= self.win_loss_ratio
self.counted_game_sets += 1
tmp = self.uncounted_games >= self.win_loss_ratio
if tmp[0] and not tmp[1]:
# Need more losses
if self.mm == 1:
# Zero weights for opponents that have <= ELOs than agent
for i, elo in enumerate(self.all_elos):
if elo <= self.agent_elo:
weights[i] = 0
# Choose agent with highest ELO if agent ELO is higher than all opponent ELOs
if sum(weights) == 0:
weights[self.all_elos.index(max(self.all_elos))] = 1
else:
# Equal probability for opponents that have > ELOs than agent
for i, elo in enumerate(self.all_elos):
if elo > self.agent_elo:
weights[i] = 1
# Choose agent with highest ELO if agent ELO is higher than all opponent ELOs
if sum(weights) == 0:
weights[self.all_elos.index(max(self.all_elos))] = 1
elif not tmp[0] and tmp[1]:
# Need more wins
if self.mm == 1:
# Zero weights for opponents that have >= ELOs than agent
for i, elo in enumerate(self.all_elos):
if elo >= self.agent_elo:
weights[i] = 0
# Choose agent with lowest ELO if agent ELO is higher than all opponent ELOs
if sum(weights) == 0:
weights[self.all_elos.index(min(self.all_elos))] = 1
else:
# Equal probability for opponents that have < ELOs than agent
for i, elo in enumerate(self.all_elos):
if elo < self.agent_elo:
weights[i] = 1
# Choose agent with highest ELO if agent ELO is higher than all opponent ELOs
if sum(weights) == 0:
weights[self.all_elos.index(min(self.all_elos))] = 1
self.current_opp_idx = choice_with_normalization(
[i for i in range(len(self.all_elos))], weights)
self.current_opp = self.all_opps[self.current_opp_idx]
self.current_opp_elo = self.all_elos[self.current_opp_idx]
#print("thread", self.my_port, "current opp elo:", self.current_opp_elo, "agent elo:", self.agent_elo, flush=True)
self.env.load_new_opp(0, opp_fp(self.model_dir, self.current_opp),
self.current_opp_elo)
def get_agent_elo(self):
return self.agent_elo
def reset(self):
if self.started:
last_winner = self.env.last_winner
if last_winner == 0:
win_rate = 1.
self.uncounted_games[0] += 1
elif last_winner == 1:
win_rate = 0.
self.uncounted_games[1] += 1
else:
win_rate = .5
agent_elo_change, _ = elo_change(self.agent_elo,
self.current_opp_elo, self.K,
win_rate)
self.agent_elo += int(agent_elo_change)
#print("THREAD", self.my_port, "CURRENT AGENT ELO:", self.agent_elo, flush=True)
else:
self.started = True
self.next_opp()
return self.env.reset()
def step(self, action):
if self.num_steps % self.elo_log_interval == 0:
self.elo_log.append(self.agent_elo)
self.num_steps += 1
return self.env.step(action)
def render(self, mode='console'):
raise NotImplementedError()
def close(self):
self.env.close()
def human_matchmaking(args):
WINS = 0
LOSSES = 1
GAMES = 2
pop = load_pop(args.model_dir)
all_stats = {}
for p in pop:
all_stats[p] = load_stats(args.model_dir, p)
all_opps = sorted_keys(all_stats)
all_opps.reverse()
all_elos = []
for opp in all_opps:
all_elos.append(int(avg_elo(all_stats[opp], avg_len=args.avg_len)))
human_stats = get_human_stats(args.human_db)
current_opp_idx = len(all_elos) // 2
current_opp = all_opps[current_opp_idx]
current_opp_elo = all_elos[current_opp_idx]
human_elo = human_stats["elo"][-1] if len(
human_stats["elo"]) > 0 else current_opp_elo
try:
env = TankEnv(args.game_path,
opp_fp_and_elo=[(opp_fp(args.model_dir,
current_opp), current_opp_elo)],
game_port=args.base_port,
my_port=args.my_port,
image_based=args.image_based,
level_path=args.level_path,
p=args.env_p)
print("Starting matchmaking")
while human_elo <= all_elos[-1]:
print("Current opp:", current_opp)
print("Opp elo:", current_opp_elo)
print("Human elo:", human_elo)
score = play_match(env, args.num_games)
human_win_rate = (
(score[WINS] - score[LOSSES]) / sum(score) + 1) / 2
K = 16
human_elo_change, _ = elo_change(human_elo, current_opp_elo, K,
human_win_rate)
human_elo += int(human_elo_change)
human_stats["elo"].append(human_elo)
if not current_opp in human_stats["win_rate"]:
human_stats["win_rate"][current_opp] = [0, 0, 0]
human_stats["win_rate"][current_opp][WINS] += score[WINS]
human_stats["win_rate"][current_opp][LOSSES] += score[LOSSES]
human_stats["win_rate"][current_opp][GAMES] += sum(score)
D = 5.
current_opp_idx = elo_based_choice(all_elos, human_elo, D)
current_opp = all_opps[current_opp_idx]
current_opp_elo = all_elos[current_opp_idx]
env.load_new_opp(0, opp_fp(args.model_dir, current_opp),
current_opp_elo)
print("CONGRATS, YOU ARE BETTER THAN ALL THE AGENTS!")
finally:
env.close()
model = PPO("CnnPolicy", env, n_steps=64)
else:
model = PPO("MlpPolicy", env, n_steps=64)
print(model.policy)
try:
if args.train:
model.learn(total_timesteps=args.num_steps)
else:
obs = env.reset()
if args.image_based and args.ai_view:
fig = plt.gcf()
fig.show()
fig.canvas.draw()
for _ in tqdm(range(args.num_steps)):
if args.image_based and args.ai_view:
plt.imshow(obs, origin="lower", interpolation='none')
fig.canvas.draw()
if model:
action, _ = model.predict(obs)
elif args.rand_p1:
action = np.random.rand(5) * 2 - 1
else:
action = np.zeros(5, dtype=np.float32)
obs, reward, done, info = env.step(action)
if done:
obs = env.reset()
finally:
env.close()
rand_opp=True)
if not args.canvas_game_path:
args.canvas_game_path = args.game_path
canvas = TankEnv(args.canvas_game_path,
opp_fp_and_elo=[],
game_port=args.base_port + 1,
my_port=args.my_port + 1,
image_based=True,
level_path=args.level_path,
rand_opp=True,
p=args.env_p)
obs = env.reset()
for i in tqdm(range(args.num_obs)):
if i % (args.num_obs // 10) == 0:
print(i / (args.num_obs // 10), "% complete", sep="", flush=True)
# Save states
obs_set[i] = obs.copy()
canvas.draw_state(obs)
img_set[i] = canvas.state.copy()
# Generate next observation
action = np.random.rand(5) * 2 - 1
obs, _, done, _ = env.step(action)
if done:
obs = env.reset()
savez_compressed(args.save_loc, obs=obs_set, img=img_set)
finally:
env.close()
canvas.close()