def objective(hypers):
tf.reset_default_graph()
cards = get_deck(num_values=args.num_values, num_suits=args.num_suits)
_, exploitabilities = nfsp(LeducNFSP(cards), hypers, max_train_steps=args.max_train_steps)
val = max(exploitabilities.values())
print("OBJECTIVE. Exploitabilities: {}, Value: {}".format(exploitabilities, val))
return val
def run_nfsp(hypers_list):
exploitabilities_list = []
for i, hypers in enumerate(hypers_list):
print("NFSP run {}".format(i))
print("Using hyperparameters: {}".format(hypers))
print("Training for {} steps".format(args.max_train_steps))
tf.reset_default_graph()
cards = get_deck(num_values=args.num_values, num_suits=args.num_suits)
_, exploitabilities = nfsp(LeducNFSP(cards), hypers, max_train_steps=args.max_train_steps)
exploitabilities_list.append(exploitabilities)
return hypers_list, exploitabilities_list
parser.add_argument(
'--num_values',
default=3,
type=int,
help='In OneCardPoker or Leduc, pass the number of cards to use.')
parser.add_argument('--num_suits',
default=2,
type=int,
help='In Leduc, pass the number of suits to use.')
args = parser.parse_args()
if args.game == 'Leduc':
print("Solving Leduc Hold'em")
cards = card.get_deck(num_values=args.num_values,
num_suits=args.num_suits)
n_game = leduc.create_neural_leduc(cards)
elif args.game == 'RockPaperScissors':
print("Solving rock paper scissors")
n_game = rock_paper_scissors.create_neural_rock_paper_scissors()
strategy, exploitabilities = deep_cfr.deep_cfr(
n_game,
num_iters=args.num_iters,
num_traversals=args.num_traversals,
advantage_maxlen=args.advantage_maxlen,
strategy_maxlen=args.strategy_maxlen,
batch_size=args.batch_size,
num_sgd_updates=args.num_sgd_updates)
exploitability = compute_exploitability(n_game.extensive_game, strategy)