def test_rescale_reward():
# tolerance
tol = 1e-14
rng = Seeder(123).rng
for _ in range(10):
# generate random MDP
S, A = 5, 2
R = rng.uniform(0.0, 1.0, (S, A))
P = rng.uniform(0.0, 1.0, (S, A, S))
for ss in range(S):
for aa in range(A):
P[ss, aa, :] /= P[ss, aa, :].sum()
env = FiniteMDP(R, P)
# test
wrapped = RescaleRewardWrapper(env, (-10, 10))
_ = wrapped.reset()
for _ in range(100):
_, reward, _, _ = wrapped.sample(
wrapped.observation_space.sample(),
wrapped.action_space.sample())
assert reward <= 10 + tol and reward >= -10 - tol
_ = wrapped.reset()
for _ in range(100):
_, reward, _, _ = wrapped.step(wrapped.action_space.sample())
assert reward <= 10 + tol and reward >= -10 - tol
def test_mbqvi(S, A):
rng = Seeder(123).rng
for sim in range(5):
# generate random MDP with deterministic transitions
R = rng.uniform(0.0, 1.0, (S, A))
P = np.zeros((S, A, S))
for ss in range(S):
for aa in range(A):
ns = rng.integers(0, S)
P[ss, aa, ns] = 1
# run MBQVI and check exactness of estimators
env = FiniteMDP(R, P)
agent = MBQVIAgent(env, n_samples=1)
agent.fit()
assert np.abs(R - agent.R_hat).max() < 1e-16
assert np.abs(P - agent.P_hat).max() < 1e-16