def test_kuhn_poker_always_pass_p0(self):
game = pyspiel.load_game("kuhn_poker")
calc = action_value_vs_best_response.Calculator(game)
(expl, avvbr, cfrp,
player_reach_probs) = calc(0, policy.FirstActionPolicy(game),
["0", "1", "2", "0pb", "1pb", "2pb"])
self.assertAlmostEqual(expl, 1.)
np.testing.assert_allclose(
avvbr,
[
# Opening bet. If we pass, we always lose (pass-pass with op's K,
# otherwise pass-bet-pass).
# If we bet, we always win (because op's best response is to pass,
# because this is an unreachable state and we break ties in favour
# of the lowest action).
[-1, 1],
[-1, 1],
[-1, 1],
# We pass, opp bets into us. This can be either J or Q (K will pass
# because of the tie-break rules).
# So we are guaranteed to be winning with Q or K.
[-1, -2], # 0pb
[-1, 2], # 1pb
[-1, 2], # 2pb
])
np.testing.assert_allclose(cfrp,
[1 / 3, 1 / 3, 1 / 3, 1 / 6, 1 / 6, 1 / 3])
np.testing.assert_allclose([1., 1., 1., 1., 1., 1.],
player_reach_probs)
def __init__(self, game):
"""Initializes a loss calculation for the given game."""
if game.num_players() != _NUM_PLAYERS:
raise ValueError("Game {} does not have {} players.".format(
game, _NUM_PLAYERS))
self.tabular_policy = policy.TabularPolicy(game)
self.q_value_calculator = action_value_vs_best_response.Calculator(game)
def test_kuhn_poker_uniform(self):
game = pyspiel.load_game("kuhn_poker")
calc = action_value_vs_best_response.Calculator(game)
expl, avvbr, cfrp = calc(0, policy.UniformRandomPolicy(game),
["0", "1", "2", "0pb", "1pb", "2pb"])
self.assertAlmostEqual(expl, 15 / 36)
np.testing.assert_allclose(
avvbr,
[
[-1.5, -2.0], # 0 (better to pass)
[-0.5, -0.5], # 1 (same)
[0.5, 1.5], # 2 (better to bet)
[-1.0, -2.0], # 0pb - losing
[-1.0, 0.0], # 1pb - best response is bet always
[-1.0, 2.0], # 2pb - winning
])
np.testing.assert_allclose(cfrp,
[1 / 3, 1 / 3, 1 / 3, 1 / 3, 1 / 3, 1 / 3])