def test_creps_variance():
x = np.zeros(1)
opt = CREPSOptimizer(x, variance=100.0, random_state=0)
opt.init(1, 1)
r = eval_loop(x, opt)
assert_less(-1e-10, r.max())
return r
def test_creps_baseline():
x = np.zeros(1)
opt = CREPSOptimizer(x, context_features="quadratic", random_state=0)
opt.init(1, 1)
r = eval_loop(x, opt, baseline_fct=lambda x: x**2)
assert_less(-1e-7, r.max())
return r
def test_creps():
x = np.zeros(1)
opt = CREPSOptimizer(x, random_state=0)
opt.init(1, 1)
r = eval_loop(x, opt)
assert_less(-1e-8, r.max())
return r
def test_record_feedbacks():
opt = CREPSOptimizer(initial_params=np.zeros(1))
ctrl = ContextualController(environment=ContextualObjectiveFunction(),
behavior_search=JustContextualOptimizer(opt),
accumulate_feedbacks=False,
record_feedbacks=True)
returns = ctrl.learn()
assert_array_equal(returns, ctrl.feedbacks_)
def test_record_outputs():
opt = CREPSOptimizer(initial_params=np.zeros(1))
ctrl = ContextualController(environment=ContextualObjectiveFunction(),
behavior_search=JustContextualOptimizer(opt),
record_outputs=True)
returns = ctrl.learn()
assert_equal(len(returns), 10)
assert_equal(np.array(ctrl.outputs_).shape, (10, 1, 0))
def test_learn_controller_cmaes_sphere():
test_contexts = np.linspace(-5, 5, 11)[:, np.newaxis]
opt = CREPSOptimizer(initial_params=np.zeros(1), random_state=0)
ctrl = ContextualController(environment=ContextualObjectiveFunction(),
behavior_search=JustContextualOptimizer(opt),
n_episodes=200,
n_episodes_before_test=200,
test_contexts=test_contexts)
ctrl.learn()
for d in ctrl.test_results_[-1]:
assert_greater(d, -1e-5)
def test_context_cannot_be_set():
class EnvironmentWithRandomContext(ContextualObjectiveFunction):
def request_context(self, _):
self.context = np.random.randn(self.n_context_dims)
return self.context
test_contexts = np.linspace(-5, 5, 11)[:, np.newaxis]
opt = CREPSOptimizer(initial_params=np.zeros(1), random_state=0)
ctrl = ContextualController(environment=EnvironmentWithRandomContext(),
behavior_search=JustContextualOptimizer(opt),
n_episodes=2,
n_episodes_before_test=1,
test_contexts=test_contexts)
assert_raises_regexp(Exception, "could not set context", ctrl.learn)
random_state = np.random.RandomState(0)
initial_params = 4.0 * np.ones(1)
n_samples_per_update = 30
variance = 0.03
context_features = "quadratic"
ccmaes = CCMAESOptimizer(initial_params=initial_params,
n_samples_per_update=n_samples_per_update,
variance=variance,
context_features=context_features,
random_state=0)
creps = CREPSOptimizer(initial_params=initial_params,
n_samples_per_update=n_samples_per_update,
train_freq=n_samples_per_update,
variance=variance,
epsilon=2.0,
context_features=context_features,
random_state=0)
opts = {"C-CMA-ES": ccmaes, "C-REPS": creps}
for opt in opts.values():
opt.init(1, 1)
n_generations = 16
n_rows = 4
params = np.empty(1)
rewards = dict([(k, []) for k in opts.keys()])
test_contexts = np.arange(-6, 6, 0.1)
colors = {"C-CMA-ES": "r", "C-REPS": "g"}
plt.figure(figsize=(n_generations * 3 / n_rows, 3 * n_rows))
for it in range(n_generations):
def test_controller_creps_objective():
opt = CREPSOptimizer(initial_params=np.zeros(1))
ctrl = ContextualController(environment=ContextualObjectiveFunction(),
behavior_search=JustContextualOptimizer(opt))
returns = ctrl.learn()
assert_equal(len(returns), 10)