def test_best_response_rps():
bart_simpson_strategy = FixedStrategy(Explicit([1, 0, 0], values=["R", "P", "S"]))
game = RockPaperScissors()
strategy = BestResponse(game, 0, {1: bart_simpson_strategy})
assert list(strategy.best_responses.values())[0].probability("R") == 0.0
assert list(strategy.best_responses.values())[0].probability("P") == 1.0
assert list(strategy.best_responses.values())[0].probability("S") == 0.0
assert strategy.value == pytest.approx(1.0)
strategy = BestResponse(game, 1, {0: bart_simpson_strategy})
assert list(strategy.best_responses.values())[0].probability("R") == 0.0
assert list(strategy.best_responses.values())[0].probability("P") == 1.0
assert list(strategy.best_responses.values())[0].probability("S") == 0.0
assert strategy.value == pytest.approx(1.0)
def test_best_response_rps():
bart_simpson_strategy = ConstStrategy((1, 0, 0))
game = RockPaperScissors()
for p in [0, 1]:
strategy = BestResponse(game, p, [bart_simpson_strategy] * 2)
assert tuple(strategy.best_responses.values())[0] == pytest.approx(
(0.0, 1.0, 0.0))
assert strategy.value == pytest.approx(1.0)
def test_best_response_goofspiel():
for n_cards, br_value in [(3, pytest.approx(4/3)), (4, pytest.approx(2.5))]:
game = Goofspiel(n_cards, Goofspiel.Scoring.ZEROSUM)
strategy = BestResponse(game, 0, {1: UniformStrategy()})
for k, v in strategy.best_responses.items():
reward = k[1][-1]
assert reward not in v.values() or v.probability(reward) == 1.0
assert strategy.value == br_value
def test_mccfr_dicepoker():
g = DicePoker()
mc = OutcomeMCCFR(g, seed=52)
mc.compute(10000, burn=0.5)
br0 = BestResponse(g, 0, mc.strategies)
assert br0.value < 0.3
payoff0 = sample_payoff(g, [br0, mc.strategies[1]], 10000, seed=3)[0]
assert br0.value == pytest.approx(payoff0[0], abs=0.05)
br1 = BestResponse(g, 1, mc.strategies)
assert br1.value > -0.2
payoff1 = sample_payoff(g, [mc.strategies[0], br1], 10000, seed=4)[0]
assert br1.value == pytest.approx(payoff1[1], abs=0.05)
print(br0.value, br1.value, payoff0, payoff1)
assert payoff0[0] > payoff1[0]
assert payoff0[0] < 0.3
assert payoff1[0] > 0.1
def test_onecardpoker_mc():
g = OneCardPoker()
mc = OutcomeMCCFR(g, seed=56)
mc.compute(1000)
#print(mc.iss)
br = BestResponse(g, 1, mc.strategies)
#print(br.value)
assert np.mean([
play_strategies(g, [mc.strategies[0], br], seed=i).payoff[0]
for i in range(1000)
]) > -0.4
def test_best_response_goofspiel():
for n_cards, br_value in [(3, pytest.approx(4 / 3)),
(4, pytest.approx(2.5))]:
game = Goofspiel(n_cards, Goofspiel.Scoring.ZEROSUM)
strategy = BestResponse(game, 0, [UniformStrategy()] * 2)
for k, v in strategy.best_responses.items():
reward = k[-1]
played_cards = k[0::3]
idx = len([
i for i in range(n_cards)
if i < reward and i not in played_cards
])
assert reward in played_cards or v[idx] == 1.0
assert strategy.value == br_value
def non_test_goofspiel():
g = Goofspiel(4, scoring=Goofspiel.Scoring.ZEROSUM)
mc = OutcomeMCCFR(g, seed=42)
for s in [10, 100, 1000]:
mc.compute(s)
br = BestResponse(g, 0, [None, mc])
print(
"Exploit after", s,
np.mean([
g.play_strategies([br, mc], seed=i)[-1].values()[0]
for i in range(1000)
]))
vs = GoofSpielCardsValueStore(g)
val = SparseStochasticValueLearning(g, vs, seed=43)
for alpha in [0.1, 0.01, 0.01, 0.001, 0.0001]:
print(alpha)
val.compute([mc, mc], 200, alpha)
def test_best_response_limit():
game = Goofspiel(3)
BestResponse(game, 0, [UniformStrategy()] * 2)
with pytest.raises(LimitExceeded, message="traversed more than"):
BestResponse(game, 0, [UniformStrategy()] * 2, max_nodes=1024)