def test_goofspiel():
g = Goofspiel(4, scoring=Goofspiel.Scoring.ZEROSUM)
mc = OutcomeMCCFR(g, seed=42)
mc.compute(100)
vs = LinearValueStore(goofspiel_feaures_cards(g.initial_state()), fix_mean=2.5)
infosampler = InformationSetSampler(g, mc)
val = SparseSGDLinearValueLearning(g, goofspiel_feaures_cards, vs, infosampler, seed=43)
val.compute([mc, mc], 100, 0.1, 0.01)
print(vs.values)
def main():
N = 4
ITERS = 2000000
g = Goofspiel(N, scoring=Goofspiel.Scoring.ZEROSUM)
mc = OutcomeMCCFR(g, seed=56)
fname = "goof-{}".format(N)
its = 1024
while its < ITERS:
cached = mc.persist(fname, iterations=its)
if not cached:
print("Exploitability after {:7d} turns (mc, g): {}, {}".format(
its, exploitability(g, 0, mc), exploitability(g, 1, mc)))
its *= 2
infosampler = InformationSetSampler(g, mc)
vsts = (1, 3)
gsts = (1, 3)
ax0 = plt.subplot(len(vsts), len(gsts), 1)
for i, (vst, gst) in enumerate(itertools.product(vsts, gsts)):
vs = LinearValueStore(goofspiel_feaures_cards(g.initial_state()),
fix_mean=(N + 1) / 2.0)
vl = SparseSGDLinearValueLearning(g,
goofspiel_feaures_cards,
vs,
infosampler,
seed=44)
vals = np.concatenate([
vl.compute([mc, mc],
1000,
step=s,
record_every=1,
val_samples=vst,
grad_samples=gst)
for s in [2**-8, 2**-9, 2**-10, 2**-11]
],
axis=0)
#c = ['red', 'green', 'blue', 'black'][i]
ax = plt.subplot(len(vsts), len(gsts), i + 1, sharex=ax0, sharey=ax0)
ax.plot(vals)
ax.legend(list(range(1, N + 1)))
ax.set_title("valseps={} gradsteps={}".format(vst, gst))
print("Done sampling valseps={} gradsteps={}".format(vst, gst))
print("Values:", vs.values)
plt.show()
return
g2 = Goofspiel(N, scoring=Goofspiel.Scoring.ZEROSUM, rewards=vs.values)
mc2 = OutcomeMCCFR(g2, seed=57)
mc2.compute(iterations=ITERS)
print("Exp(mc2, g2)", exploitability(g2, 0, mc2),
exploitability(g2, 1, mc2))
print("Exp(mc2, g)", exploitability(g, 0, mc2), exploitability(g, 1, mc2))
def test_goofspeil():
g = Goofspiel(7)
s = g.initial_state()
assert s.player() == s.P_CHANCE
assert s.score(0) == 0
assert s.score(1) == 0
assert s.actions() == list(range(7))
assert (
s.chance_distribution().probabilities() == (pytest.approx(1 / 7), ) *
7).all()
for i, a in enumerate([3, 1, 0, 4, 3, 5, 5, 2, 2, 1, 4, 3, 2, 6]):
s = s.play(a)
assert s.player() == (i + 1) % 3 - 1
assert s.round() == 4
assert s.player() == 1
assert s.actions() == [1, 4, 6]
assert s.winners() == [0, 1, -1, 0]
assert (
s.chance_distribution().probabilities() == (pytest.approx(1.0 / 3), ) *
3).all()
assert s.score(0) == 6
assert s.score(1) == 5
assert s.cards_in_hand(-1) == [0, 6]
assert s.cards_in_hand(0) == [0, 5]
assert s.cards_in_hand(1) == [1, 4, 6]
for a in [1, 6, 5, 6, 0, 0, 0]:
s = s.play(a)
assert s.is_terminal()
assert s.score(0) == 9
assert s.score(1) == 12
assert s.values() == (-1, 1)