def test_continuous_state_env_2():
env = MountainCar()
counter = OnlineDiscretizationCounter(env.observation_space,
env.action_space)
for N in [10, 20, 30]:
for _ in range(100):
ss = env.observation_space.sample()
aa = env.action_space.sample()
for nn in range(N):
ns, rr, _, _ = env.sample(ss, aa)
counter.update(ss, aa, ns, rr)
assert counter.count(ss, aa) == N
counter.reset()
def test_continuous_state_env():
env = MountainCar()
counter = DiscreteCounter(env.observation_space, env.action_space)
for N in [10, 20, 30]:
for _ in range(100):
ss = env.observation_space.sample()
aa = env.action_space.sample()
for _ in range(N):
ns, rr, _, _ = env.sample(ss, aa)
counter.update(ss, aa, ns, rr)
dss = counter.state_discretizer.discretize(ss)
assert counter.N_sa[dss, aa] == N
assert counter.count(ss, aa) == N
counter.reset()
def test_continuous_state_env_2(rate_power):
env = MountainCar()
counter = OnlineDiscretizationCounter(env.observation_space,
env.action_space,
rate_power=rate_power)
for N in [10, 20]:
for _ in range(50):
ss = env.observation_space.sample()
aa = env.action_space.sample()
for nn in range(N):
ns, rr, _, _ = env.sample(ss, aa)
counter.update(ss, aa, ns, rr)
assert counter.count(ss, aa) == N
if rate_power == pytest.approx(1):
assert np.allclose(counter.measure(ss, aa), 1.0 / N)
elif rate_power == pytest.approx(0.5):
assert np.allclose(counter.measure(ss, aa), np.sqrt(1.0 / N))
counter.reset()
def test_dqn_agent_rnd():
env = MountainCar()
def uncertainty_estimator_fn(observation_space, action_space):
counter = RandomNetworkDistillation(observation_space, action_space)
return counter
agent = DQNAgent(env,
n_episodes=2,
use_bonus=True,
uncertainty_estimator_kwargs=dict(
uncertainty_estimator_fn=uncertainty_estimator_fn,
bonus_scale_factor=1.0))
agent.fit()
agent.policy(env.observation_space.sample())
def test_dqn_agent():
env = MountainCar()
def uncertainty_estimator_fn(observation_space, action_space):
counter = OnlineDiscretizationCounter(observation_space,
action_space,
min_dist=0.25)
return counter
agent = DQNAgent(env,
n_episodes=2,
use_bonus=True,
uncertainty_estimator_kwargs=dict(
uncertainty_estimator_fn=uncertainty_estimator_fn,
bonus_scale_factor=1.0))
agent.fit()
agent.policy(env.observation_space.sample())