def test_solver_leduc_poker_first_10_steps_convergence(self):
solver = exploitability_descent.Solver(
pyspiel.load_game("leduc_poker"))
with self.session() as session:
session.run(tf.global_variables_initializer())
nash_conv = [
solver.step(session, learning_rate=1.0) for _ in range(11)
]
# Nash conv is 2x exploitability. Values taken from test run, not
# independently verified.
np.testing.assert_allclose(nash_conv, [
4.7472224, 4.3147216, 3.9900389, 3.7576618, 3.5771275, 3.4414644,
3.3272073, 3.1898201, 3.1089299, 3.0108435, 2.8992782
])
def test_solver_kuhn_poker_first_10_steps_convergence(self):
solver = exploitability_descent.Solver(pyspiel.load_game("kuhn_poker"))
with self.session() as session:
session.run(tf.global_variables_initializer())
nash_conv = [
solver.step(session, learning_rate=1.0) for _ in range(11)
]
# Nash conv is 2x exploitability. Values taken from test run, not
# independently verified.
np.testing.assert_allclose(nash_conv, [
0.91666666666666652, 0.67893004801213452, 0.48109148836354743,
0.40061420923255808, 0.36617242161468722, 0.33676996443499557,
0.30925081512398128, 0.28827843035940964, 0.26830042206858751,
0.24418597846799289, 0.22168699344791482
])
def test_solver_kuhn_poker_first_10_steps_convergence(self):
solver = exploitability_descent.Solver(pyspiel.load_game("kuhn_poker"))
with self.session() as session:
session.run(tf.global_variables_initializer())
nash_conv = [
solver.step(session, learning_rate=1.0) for _ in range(11)
]
# Nash conv is 2x exploitability. Reference exploitability values here:
# google3/third_party/open_spiel/python/algorithms/exploitability_descent_test.py?rcl=235690821
np.testing.assert_allclose(nash_conv, [
0.91666666666666652, 0.67893004801213452, 0.48109148836354743,
0.40061420923255808, 0.36617242161468722, 0.33676996443499557,
0.30925081512398128, 0.28827843035940964, 0.26830042206858751,
0.24418597846799289, 0.22168699344791482
])