def test_identity_multidiscrete(model_class):
"""
Test if the algorithm (with a given policy)
can learn an identity transformation (i.e. return observation as an action)
with a multidiscrete action space
:param model_class: (BaseRLModel) A RL Model
"""
env = DummyVecEnv([lambda: IdentityEnvMultiDiscrete(10)])
model = model_class("MlpPolicy", env)
model.learn(total_timesteps=1000, seed=0)
n_trials = 1000
reward_sum = 0
obs = env.reset()
for _ in range(n_trials):
action, _ = model.predict(obs)
obs, reward, _, _ = env.step(action)
reward_sum += reward
assert np.array(model.action_probability(obs)).shape == (2, 1, 10), \
"Error: action_probability not returning correct shape"
assert np.prod(model.action_probability(obs, actions=env.action_space.sample()).shape) == 1, \
"Error: not scalar probability"
def test_identity_multi_discrete(model_class):
"""
test the MultiDiscrete environment vectorisation detection
:param model_class: (BaseRLModel) the RL model
"""
check_shape(lambda: IdentityEnvMultiDiscrete(dim=10), model_class, (2, ),
(1, 2))
def test_identity_multi_discrete(model_class):
"""
test the MultiDiscrete environment vectorisation detection
:param model_class: (BaseRLModel) the RL model
"""
model = model_class(policy="MlpPolicy", env=DummyVecEnv([lambda: IdentityEnvMultiDiscrete(dim=10)]))
env0 = IdentityEnvMultiDiscrete(dim=10)
env1 = DummyVecEnv([lambda: IdentityEnvMultiDiscrete(dim=10)])
n_trials = 100
for env, expected_shape in [(env0, (2,)), (env1, (1, 2))]:
obs = env.reset()
for _ in range(n_trials):
action, _ = model.predict(obs)
assert np.array(action).shape == expected_shape
obs, _, _, _ = env.step(action)
# Free memory
del model, env
def test_identity_multidiscrete(model_class):
"""
Test if the algorithm (with a given policy)
can learn an identity transformation (i.e. return observation as an action)
with a multidiscrete action space
:param model_class: (BaseRLModel) A RL Model
"""
env = DummyVecEnv([lambda: IdentityEnvMultiDiscrete(10)])
model = model_class("MlpPolicy", env)
model.learn(total_timesteps=1000)
evaluate_policy(model, env, n_eval_episodes=5)
obs = env.reset()
assert np.array(model.action_probability(obs)).shape == (2, 1, 10), \
"Error: action_probability not returning correct shape"
assert np.prod(model.action_probability(obs, actions=env.action_space.sample()).shape) == 1, \
"Error: not scalar probability"
def test_identity_multidiscrete(model_class):
"""
Test if the algorithm (with a given policy)
can learn an identity transformation (i.e. return observation as an action)
with a multidiscrete action space
:param model_class: (BaseRLModel) A RL Model
"""
env = DummyVecEnv([lambda: IdentityEnvMultiDiscrete(10)])
model = model_class("MlpPolicy", env)
model.learn(total_timesteps=1000, seed=0)
n_trials = 1000
reward_sum = 0
obs = env.reset()
for _ in range(n_trials):
action, _ = model.predict(obs)
obs, reward, _, _ = env.step(action)
reward_sum += reward