def copy_strategy(dst, src):
def on_node(dst_node, src_node):
if isinstance(dst_node, ActionNode):
np.copyto(dst_node.strategy, src_node.strategy)
return [src_node.children[a] for a in src_node.children]
walk_trees(on_node, dst, src)
def read_strategy_from_file(game, strategy_file_path):
strategy = {}
with open(strategy_file_path, 'r') as strategy_file:
for line in strategy_file:
if not line.strip() or line.strip().startswith('#'):
continue
line_split = line.split(' ')
strategy[line_split[0]] = [
float(probStr) for probStr in line_split[1:4]
]
if not game:
return strategy
game_instance = acpc.read_game_file(game) if isinstance(game,
str) else game
strategy_tree = GameTreeBuilder(game_instance,
StrategyTreeNodeProvider()).build_tree()
def on_node(node):
if isinstance(node, ActionNode):
nonlocal strategy
node_strategy = np.array(strategy[str(node)])
np.copyto(node.strategy, node_strategy)
walk_trees(on_node, strategy_tree)
return strategy_tree, strategy
def is_strategies_equal(first, second):
equal = True
def on_node(first_node, second_node):
if isinstance(first_node, ActionNode):
for a in range(3):
if not isclose(first_node.strategy[a],
second_node.strategy[a]):
nonlocal equal
equal = False
walk_trees(on_node, first, second)
return equal
def __init__(self, game, strategies):
self.strategy = GameTreeBuilder(
game, StrategiesWeightedMixtureTreeNodeProvider()).build_tree()
self.weights = np.ones(len(strategies)) / len(strategies)
def on_nodes(*nodes):
mixture_node = nodes[0]
if isinstance(mixture_node, ActionNode):
mixture_node.weights = self.weights
mixture_node.strategies = np.zeros([len(strategies), 3])
for i, node in enumerate(nodes[1:]):
mixture_node.strategies[i, :] = node.strategy
walk_trees(on_nodes, self.strategy, *strategies)
def is_correct_strategy(strategy_tree):
correct = True
def on_node(node):
if isinstance(node, ActionNode):
nonlocal correct
strategy_sum = np.sum(node.strategy)
if not isclose(strategy_sum, 1):
correct = False
for i in range(3):
if i not in node.children and node.strategy[i] != 0:
correct = False
walk_trees(on_node, strategy_tree)
return correct
def test_strategy_writing_and_reading(self):
game = acpc.read_game_file(KUHN_POKER_GAME_FILE_PATH)
strategy_tree = GameTreeBuilder(game, StrategyTreeNodeProvider()).build_tree()
def on_node(node):
if isinstance(node, ActionNode):
for a in range(3):
if a in node.children:
node.strategy[a] = 0.5
else:
node.strategy[a] = 7
walk_trees(on_node, strategy_tree)
write_strategy_to_file(strategy_tree, 'test/io_test_dummy.strategy')
read_strategy_tree, _ = read_strategy_from_file(KUHN_POKER_GAME_FILE_PATH, 'test/io_test_dummy.strategy')
self.assertTrue(is_strategies_equal(strategy_tree, read_strategy_tree))
def __init__(
self,
game,
opponent_strategy_tree,
p,
show_progress=True):
super().__init__(game, show_progress)
self.p = p
opponent_strategy = {}
def callback(node):
if isinstance(node, ActionNode):
nonlocal opponent_strategy
opponent_strategy[str(node)] = node.strategy
walk_trees(callback, opponent_strategy_tree)
self.opponent_strategy = opponent_strategy
def __init__(self,
game,
opponent_sample_tree,
p_max=0.8,
show_progress=True):
super().__init__(game, show_progress)
self.p_max = p_max
opponent_action_decision_counts = {}
def callback(node):
if isinstance(node, ActionNode):
nonlocal opponent_action_decision_counts
opponent_action_decision_counts[str(
node)] = node.action_decision_counts
walk_trees(callback, opponent_sample_tree)
self.opponent_action_decision_counts = opponent_action_decision_counts
def test_leduc_rnr_works(self):
game = acpc.read_game_file(LEDUC_POKER_GAME_FILE_PATH)
opponent_strategy = GameTreeBuilder(
game, StrategyTreeNodeProvider()).build_tree()
def on_node(node):
if isinstance(node, ActionNode):
action_count = len(node.children)
action_probability = 1 / action_count
for a in node.children:
node.strategy[a] = action_probability
walk_trees(on_node, opponent_strategy)
rnr = RestrictedNashResponse(game,
opponent_strategy,
0.5,
show_progress=False)
rnr.train(10, 5)
def create_agent_strategy_from_trained_strategy(
game_file_path,
strategy_tree,
tilt_action,
tilt_type,
tilt_probability,
in_place=False):
tilt_action_index = tilt_action.value
def on_node(node):
if tilt_action_index in node.children:
original_tilt_action_probability = node.strategy[tilt_action_index]
new_tilt_action_probability = None
if tilt_type == TiltType.ADD:
new_tilt_action_probability = np.clip(original_tilt_action_probability + tilt_probability, 0, 1)
elif tilt_type == TiltType.MULTIPLY:
new_tilt_action_probability = np.clip(
original_tilt_action_probability + original_tilt_action_probability * tilt_probability, 0, 1)
node.strategy[tilt_action_index] = new_tilt_action_probability
diff = new_tilt_action_probability - original_tilt_action_probability
other_actions_probability = 1 - original_tilt_action_probability
if diff != 0 and other_actions_probability == 0:
other_action_probability_diff = diff / (len(node.children) - 1)
for a in filter(lambda a: a != tilt_action_index, node.children):
node.strategy[a] -= other_action_probability_diff
elif diff != 0:
for a in filter(lambda a: a != tilt_action_index, node.children):
node.strategy[a] -= diff * (node.strategy[a] / other_actions_probability)
result_strategy = None
if in_place:
result_strategy = strategy_tree
else:
game = acpc.read_game_file(game_file_path)
result_strategy = GameTreeBuilder(game, StrategyTreeNodeProvider()).build_tree()
copy_strategy(result_strategy, strategy_tree)
walk_trees(on_node, result_strategy)
return result_strategy
def create_strategy(self, game, node_strategy_creator_callback):
strategy = GameTreeBuilder(game,
StrategyTreeNodeProvider()).build_tree()
walk_trees(node_strategy_creator_callback, strategy)
return strategy
def train_and_show_results(self, test_spec):
game = acpc.read_game_file(test_spec['game_file_path'])
weak_opponent_samples_tree = GameTreeBuilder(
game, SamplesTreeNodeProvider()).build_tree()
weak_opponent_strategy_tree = GameTreeBuilder(
game, StrategyTreeNodeProvider()).build_tree()
def on_node(samples_node, strategy_node):
if isinstance(samples_node, ActionNode):
child_count = len(samples_node.children)
samples_count = random.randrange(15)
for i, a in enumerate(samples_node.children):
if i < (child_count - 1) and samples_count > 0:
action_samples_count = random.randrange(samples_count +
1)
samples_count -= action_samples_count
samples_node.action_decision_counts[
a] = action_samples_count
else:
samples_node.action_decision_counts[a] = samples_count
samples_sum = np.sum(samples_node.action_decision_counts)
if samples_sum > 0:
strategy_node.strategy = samples_node.action_decision_counts / samples_sum
else:
for a in strategy_node.children:
strategy_node.strategy[a] = 1 / len(
strategy_node.children)
walk_trees(on_node, weak_opponent_samples_tree,
weak_opponent_strategy_tree)
self.assertTrue(is_correct_strategy(weak_opponent_strategy_tree))
exploitability = Exploitability(game)
num_test_counts = test_spec['test_counts']
data = np.zeros([num_test_counts, 2, len(P_MAX_VALUES)])
for i in range(num_test_counts):
print('%s/%s' % (i + 1, num_test_counts))
for j, p_max in enumerate(P_MAX_VALUES):
print('Pmax: %s - %s/%s' % (p_max, j + 1, len(P_MAX_VALUES)))
dbr = DataBiasedResponse(game,
weak_opponent_samples_tree,
p_max=p_max)
dbr.train(test_spec['training_iterations'])
data[i, 0, j] = exploitability.evaluate(dbr.game_tree)
data[i, 1,
j] = exploitability.evaluate(weak_opponent_strategy_tree,
dbr.game_tree)
plt.figure(dpi=160)
for k in range(i + 1):
run_index = math.floor(k / 2)
xdata = data[k,
0, :] if k < i or j == (len(P_MAX_VALUES) -
1) else data[k, 0,
0:j + 1]
ydata = data[k,
1, :] if k < i or j == (len(P_MAX_VALUES) -
1) else data[k, 1,
0:j + 1]
plt.plot(xdata,
ydata,
label='Run %s' % (run_index + 1),
marker='o',
linewidth=0.8)
if 'title' in test_spec:
plt.title(test_spec['title'])
plt.xlabel('DBR trained strategy exploitability [mbb/g]')
plt.ylabel(
'Random opponent exploitation by DBR strategy [mbb/g]')
plt.grid()
if num_test_counts > 1:
plt.legend()
game_name = test_spec['game_file_path'].split('/')[1][:-5]
figure_output_path = '%s/%s(it:%s).png' % (
FIGURES_FOLDER, game_name,
test_spec['training_iterations'])
figures_directory = os.path.dirname(figure_output_path)
if not os.path.exists(figures_directory):
os.makedirs(figures_directory)
plt.savefig(figure_output_path)
print('\033[91mThis test needs your assistance! ' +
'Check the generated graph %s!\033[0m' % figure_output_path)
