def on_start(passed_room=None):
if passed_room is not None:
room = passed_room
else:
room = rt.sids_to_rooms[request.sid]
if not rt.game_rooms[room]["started"]: # Don't allow multiple starts
print("Starting")
broadcast_to_room(room)("", "start game")
rt.game_rooms[room]["started"] = True
# shuffle clients randomly
clients_keys = list(rt.game_rooms[room]["clients"].keys())
random_keys = [i for i in range(len(rt.game_rooms[room]["clients"]))]
if not keep_client_order:
random.shuffle(random_keys)
shuffled_clients = {}
for i, k in enumerate(random_keys):
shuffled_clients[k] = rt.game_rooms[room]["clients"][clients_keys[i]]
rt.game_rooms[room]["clients"] = shuffled_clients
ai_players = [c for c in rt.game_rooms[room]["clients"] if rt.game_rooms[room]["clients"][c]["ai"]]
print(rt.game_rooms[room]["clients"])
engine = Engine(emit_to_client_in_room(room), broadcast_to_room(room), retrieve_response_in_room(room),
nonlocal_ais=ai_players, n_players=len(rt.game_rooms[room]["clients"]), **parsed_args)
winner = engine.run_game()
socketio.stop()
def benchmark(self, agent: Agent, n=100):
stats = {}
# bench marking against other Agents
opponents = [
TaxAgent(),
IncomeAgent(),
AdversarialAgent(),
HonestAgent(),
StrategicAgentV1(),
TrickyAgent()
]
# remove epsilon for benchmarking
agent.epsilon = 0
agent.verbose = False
for opponent in opponents:
wins = 0
opponent_name = str(opponent)
print(f"Benchmarking against {opponent_name}...")
for i in range(n):
engine = Engine(local_ais={0: agent, 1: opponent})
winner = engine.run_game()
wins += 1 if winner == 0 else 0
win_rate = wins / (i + 1)
print("Win rate: ", win_rate)
stats[opponent_name] = win_rate
print(stats)
return stats
def evaluate_with_model(self, training_location, eval_model_location, n):
wins = 0
for i in range(n):
engine = Engine(local_ais={0: IncomeAgent(), 1: IncomeAgent()})
winner = engine.run_game()
wins += 1 if winner == 0 else 0
print("WIN RATE: ", wins / (i + 1))
return wins / n
def self_train(self, training_location, hold_location):
n = 10000
wins = 0
for i in range(n):
print("TRAINING EPISODE: ", i)
if (i % 1000 == 0):
self.update_hold(hold_location, training_location)
engine = Engine(local_ais={0: KerasAgent(), 1: KerasAgent()})
winner = engine.run_game()
wins += 1 if winner == 0 else 0
print("WIN RATE: ", wins / (i + 1))
return wins / n
def train_epoch(self, training_agent: Agent, op_agent: Agent):
"""
prints(win rate, total reward, loss)
returns xs, rewards
"""
for _ in range(self.epoch):
engine = Engine(local_ais={0: training_agent, 1: op_agent})
winner = engine.run_game()
xs, ys, rewards, game_histories = training_agent.get_training_data()
xs = np.vstack(xs)
discounted = np.vstack(self.discount_rewards(rewards))
training_history = training_agent.model.fit(xs, discounted)
# reset stats
training_agent.reset()
return training_history, game_histories
def run_match(agents,
elos=None,
randomize_order=True,
n_iters=1000,
use_default_elos=False):
if elos is None:
if use_default_elos:
elos = [DEFAULT_ELO] * len(agents)
else:
raise ValueError(
"Must specify initial elos or specify to use default elos")
old_elos = [elo for elo in elos]
print("Agents:", str([str(agent) for agent in agents]).replace("'", ""))
print("Old elos:", [round(elo) for elo in old_elos])
agent_order = [i for i in range(len(agents))]
win_tracker = {}
for i in range(len(agents)):
win_tracker[i] = 0
for i in range(n_iters):
if randomize_order:
random.shuffle(agent_order)
local_ais = {}
for j in range(len(agents)):
local_ais[j] = agents[agent_order[j]]
engine = Engine(local_ais=local_ais)
winner = engine.run_game()
win_tracker[agent_order[winner]] += 1
if (i + 1) % 100 == 0:
print(f"Win proportions after game {i+1}:", [
round(win_tracker[j] / (i + 1), 3)
for j in range(len(win_tracker))
])
elos = calculate_new_elos(elos, agent_order[winner])
print("Final win proportions:",
[win_tracker[i] / n_iters for i in range(len(win_tracker))])
print("New elos:", [round(elo) for elo in elos])
elo_diffs = [elos[i] - old_elos[i] for i in range(len(elos))]
elo_diffs = [("+" if diff >= 0 else "") + str(round(diff))
for diff in elo_diffs]
print("Elo changes:", str(elo_diffs).replace("'", ""))
return elos
local_ais = {0: a, 1: b}
# print(a.model.hidden_state)
# print(b.model.hidden_state)
# for n, p in a.model.named_parameters():
# print(n, p.data)
print("Starting Training")
training_winners = []
n_train_iters = 10000
for i in range(n_train_iters):
if i % 100 == 0:
print("Training Iteration " + str(i + 1))
print("Win percentage this epoch:",
len([r for r in training_winners[i - 100:i] if r == 0]))
print([x for x in a.model.named_parameters()][-1])
e = Engine(local_ais=local_ais)
winner = e.run_game()
training_winners.append(winner)
# print(a.model.hidden_state)
# print(b.model.hidden_state)
# time.sleep(2)
# for name, p in a.model.named_parameters():
# print(name)
# print(p.data)
# time.sleep(2)
# break
a.train_model(winner)
a.events = []
# for name, p in a.model.named_parameters():