def test_expected_utility_three_player():
num_players = 3
node_map = {i: {} for i in range(num_players)}
action_map = {i: {} for i in range(num_players)}
cards = [Card(14, 1), Card(13, 1), Card(12, 1), Card(11, 1)]
learn(20000, cards, 3, node_map, action_map)
util = expected_utility(cards, 3, 3, kuhn_eval, node_map, action_map)
print(util)
print(json.dumps(action_map, indent=4))
print(node_map)
assert abs(util.sum()) <= 0.0001, f"Something weird, not a zero sum game"
assert np.abs(util).sum() > 0, f"Util was {util}"
def test_expected_utility():
num_players = 2
node_map = {i: {} for i in range(num_players)}
action_map = {i: {} for i in range(num_players)}
cards = [Card(14, 1), Card(13, 1), Card(12, 1)]
learn(20000, cards, 2, node_map, action_map)
util = expected_utility(cards, 2, 2, kuhn_eval, node_map, action_map)
print(util)
print(json.dumps(action_map, indent=4))
print(node_map)
assert util.sum() == 0
assert np.isclose(np.abs(util[0]), np.abs(util[1])), f"Util was {util}"
assert np.abs(util).sum() > 0, f"Util was {util}"
assert abs(util[1] - 1 / 18) <= .01, f"Util not converging {util}"
def learn(iterations, node_map, action_map):
num_players = len(node_map)
cards = [Card(14 - i, 1) for i in range(num_players + 1)]
for i in tqdm(range(iterations), desc="learning"):
np.random.shuffle(cards)
for player in range(num_players):
state = State(cards, num_players, 1, kuhn_eval)
probs = np.ones(num_players)
accumulate_regrets(state, node_map, action_map, probs)
def test_update_strategy():
num_players = 2
node_map = {i: {} for i in range(num_players)}
action_map = {i: {} for i in range(num_players)}
n1 = Node(['F', 'C', '1R'])
n1.regret_sum = {'F': 0, 'C': 1, '1R': 0}
n2 = Node(['F', 'C', '1R'])
n2.regret_sum = {'F': 1, 'C': 0, '1R': 1}
node_map[0]['As || [[]]'] = n1
node_map[0]["As || [['C', '1R']]"] = n2
cards = [Card(14, 1), Card(13, 1), Card(12, 1)]
state = State(cards, num_players, 1, kuhn_eval)
update_strategy(0, state, node_map, action_map)
assert sum(n1.strategy_sum.values()) > 0, f'Util\n{n1}, \nNodes\n{node_map}, Actions\n{json.dumps(action_map, indent=4)}'
assert sum(n2.strategy_sum.values()) > 0, f'Util\n{n1}, \nNodes\n{node_map}, Actions\n{json.dumps(action_map, indent=4)}'
def test_expected_utility():
num_players = 2
node_map = {i: {} for i in range(num_players)}
action_map = {i: {} for i in range(num_players)}
cards = [Card(14, 1), Card(13, 1), Card(12, 1)]
learn(10000, cards, 2, node_map, action_map)
exploit = exploitability(cards, 2, node_map, action_map)
print(exploit)
assert exploit < .001, f"Exploitability was : {exploit}"
util = expected_utility(cards, 2, 2, node_map, action_map)
print(util)
print(json.dumps(action_map, indent=4))
print(node_map)
assert util.sum() == 0
assert np.isclose(np.abs(util[0]), np.abs(util[1])), f"Util was {util}"
assert np.abs(util).sum() > 0, f"Util was {util}"
def test_leduc_showdown():
cards = [
Card(14, 1),
Card(13, 1),
Card(12, 1),
Card(14, 2),
Card(13, 2),
Card(12, 2)
]
state = Leduc(cards, 2, leduc_eval)
state = state.take('2R', deep=True)
state = state.take('2R', deep=True)
state = state.take('C', deep=True)
assert state.round == 1 and state.turn == 0, f'{state.round, state.turn}'
state = state.take('4R', deep=True)
assert state.turn == 1, state.turn
state = state.take('C', deep=True)
assert state.terminal is True and np.array_equal(
state.utility(), np.array([9, -9])), f'{state.utility(), state.cards}'
def test_leduc_state():
cards = [
Card(14, 1),
Card(13, 1),
Card(12, 1),
Card(14, 2),
Card(13, 2),
Card(12, 2)
]
state = Leduc(cards, 2, leduc_eval)
assert state.turn == 0, state.turn
state = state.take('C', deep=True)
assert state.turn == 1 and state.round == 0, f'{state.turn, state.round}'
state = state.take('2R', deep=True)
assert state.turn == 0 and state.round == 0, f'{state.turn, state.round}'
state = state.take('C', deep=True)
assert state.turn == 0 and state.round == 1, f'{state.turn, state.round}'
state = state.take('F', deep=True)
assert state.terminal is True and np.array_equal(
state.utility(), np.array([-3, 3
])), f'{state.terminal, state.utility()}'
def test_hand_eval():
cards = [
Card(14, 1),
Card(13, 1),
Card(12, 1),
Card(14, 2),
Card(13, 2),
Card(12, 2)
]
score = leduc_eval(cards[0], [cards[3]])
assert score == 224
def test_kuhn_utility():
cards = [Card(14, 1), Card(13, 1), Card(12, 1)]
state = State(cards, 2, kuhn_eval)
state.take('C')
state.take('C')
utility = state.utility()
assert np.array_equal(utility, np.array([1, -1])), utility
cards = [Card(14, 1), Card(13, 1), Card(12, 1)]
state = State(cards, 2, kuhn_eval)
state.take('1R')
state.take('C')
utility = state.utility()
assert np.array_equal(utility, np.array([2, -2])), utility
node = node_map[hand.turn][info_set]
strategy = node.avg_strategy()
util = np.zeros(len(node_map))
valid_actions = action_map[hand.turn][info_set]
for action in valid_actions:
new_hand = hand.take(action, deep=True)
util += traverse_tree(new_hand, node_map, action_map) * strategy[action]
return util
if __name__ == '__main__':
if len(sys.argv) > 1:
try:
num_players = int(sys.argv[1])
except ValueError:
raise ValueError("must pass an int for number of players")
else:
num_players = 2
node_map = {i: {} for i in range(num_players)}
action_map = {i: {} for i in range(num_players)}
learn(10000, node_map, action_map)
cards = [Card(14, 1), Card(13, 1), Card(12, 1)]
util = expected_utility(cards, 2, 2, kuhn_eval, node_map, action_map)
print(util)
print(node_map)
print(json.dumps(action_map, indent=4))
if p != state.turn:
reach_prob *= prob
for action in valid_actions:
regret = util[action] - node_util[state.turn]
node.regret_sum[action] += regret * reach_prob
return node_util
if __name__ == '__main__':
num_players = 2
node_map = {i: {} for i in range(num_players)}
action_map = {i: {} for i in range(num_players)}
cards = [
Card(14, 1),
Card(13, 1),
Card(12, 1),
Card(14, 2),
Card(13, 2),
Card(12, 2)
]
learn(10000, cards, 3, node_map, action_map)
exploit = exploitability(cards, 3, node_map, action_map)
print(exploit)
for player in node_map:
print(f"Player {player}")
print('Number of info sets', len(node_map[player]))
for info_set, node in node_map[player].items():
avg_strat = node.avg_strategy()
util = np.zeros(len(node_map))
valid_actions = action_map[hand.turn][info_set]
for action in valid_actions:
new_hand = hand.take(action, deep=True)
util += traverse_tree(new_hand, node_map,
action_map) * strategy[action]
return util
if __name__ == '__main__':
if len(sys.argv) > 1:
try:
num_players = int(sys.argv[1])
except ValueError:
raise ValueError("must pass an int for number of players")
else:
num_players = 2
node_map = {i: {} for i in range(num_players)}
action_map = {i: {} for i in range(num_players)}
cards = [Card(14 - i, 1) for i in range(num_players + 1)]
learn(10000, cards, num_players, node_map, action_map)
exploit = exploitability(cards, 2, node_map, action_map)
print(exploit)
util = expected_utility(cards, 2, 2, node_map, action_map)
print(util)
print(node_map)
print(json.dumps(action_map, indent=4))
return node_util
else:
actions = list(strategy.keys())
probs = list(strategy.values())
random_action = actions[np.random.choice(len(actions), p=probs)]
new_state = state.take(random_action, deep=True)
return accumulate_regrets(traverser, new_state, node_map, action_map,
prune=prune)
if __name__ == '__main__':
num_players = 2
node_map = {i: {} for i in range(num_players)}
action_map = {i: {} for i in range(num_players)}
cards = [Card(14, 1), Card(13, 1), Card(12, 1), Card(14, 2), Card(13, 2), Card(12, 2)]
learn(50000, cards, 3, node_map, action_map)
exploit = exploitability(cards, 3, node_map, action_map)
print(exploit)
for player in node_map:
print(f"Player {player}")
print('Number of info sets', len(node_map[player]))
for info_set, node in node_map[player].items():
avg_strat = node.avg_strategy()
print(f"{info_set}: {avg_strat}")
util = expected_utility(cards, 3, 2, node_map, action_map)
print(util)