def test_act_noise_simple(env):
# Typical case with zero mean and non-zero std
wrapped_env = GaussianActNoiseWrapper(env, noise_std=0.2*np.ones(env.act_space.shape))
for _ in range(3):
# Sample some values
rand_act = env.act_space.sample_uniform()
wrapped_env.reset()
obs_nom, _, _, _ = env.step(rand_act)
obs_wrapped, _, _, _ = wrapped_env.step(rand_act)
# Different actions can not lead to the same observation
assert not np.all(obs_nom == obs_wrapped)
# Unusual case with non-zero mean and zero std
wrapped_env = GaussianActNoiseWrapper(env, noise_mean=0.1*np.ones(env.act_space.shape))
for _ in range(3):
# Sample some values
rand_act = env.act_space.sample_uniform()
wrapped_env.reset()
obs_nom, _, _, _ = env.step(rand_act)
obs_wrapped, _, _, _ = wrapped_env.step(rand_act)
# Different actions can not lead to the same observation
assert not np.all(obs_nom == obs_wrapped)
# General case with non-zero mean and non-zero std
wrapped_env = GaussianActNoiseWrapper(env,
noise_mean=0.1*np.ones(env.act_space.shape),
noise_std=0.2*np.ones(env.act_space.shape))
for _ in range(3):
# Sample some values
rand_act = env.act_space.sample_uniform()
wrapped_env.reset()
obs_nom, _, _, _ = env.step(rand_act)
obs_wrapped, _, _, _ = wrapped_env.step(rand_act)
# Different actions can not lead to the same observation
assert not np.all(obs_nom == obs_wrapped)
def test_order_act_noise_act_norm(env: SimEnv):
# First noise wrapper then normalization wrapper
wrapped_env_noise = GaussianActNoiseWrapper(
env,
noise_mean=0.2 * np.ones(env.act_space.shape),
noise_std=0.1 * np.ones(env.act_space.shape))
wrapped_env_noise_norm = ActNormWrapper(wrapped_env_noise)
# First normalization wrapper then noise wrapper
wrapped_env_norm = ActNormWrapper(env)
wrapped_env_norm_noise = GaussianActNoiseWrapper(
wrapped_env_norm,
noise_mean=0.2 * np.ones(env.act_space.shape),
noise_std=0.1 * np.ones(env.act_space.shape))
# Sample some values directly from the act_spaces
for i in range(3):
pyrado.set_seed(i)
act_noise_norm = wrapped_env_noise_norm.act_space.sample_uniform()
pyrado.set_seed(i)
act_norm_noise = wrapped_env_norm_noise.act_space.sample_uniform()
# These samples must be the same since were not passed to _process_act function
assert np.allclose(act_noise_norm, act_norm_noise)
# Process a sampled action
for i in range(3):
# Sample a small random action such that the de-normalization does not map it to the act_space limits
rand_act = 0.01 * env.act_space.sample_uniform()
pyrado.set_seed(i)
wrapped_env_noise_norm.reset()
obs_noise_norm, _, _, _ = wrapped_env_noise_norm.step(rand_act)
pyrado.set_seed(i)
wrapped_env_norm_noise.reset()
obs_norm_noise, _, _, _ = wrapped_env_norm_noise.step(rand_act)
# The order of processing (first normalization or first randomization must make a difference)
assert not np.allclose(obs_noise_norm, obs_norm_noise)