def test_noncontextual_behavior_search():
opt = CMAESOptimizer(initial_params=np.zeros(1))
assert_raises_regexp(TypeError,
"requires contextual behavior search",
ContextualController,
environment=ContextualObjectiveFunction(),
behavior_search=JustOptimizer(opt))
def test_set_context():
env = ObjectiveFunction(name="Sphere", n_params=2, random_state=0)
env.init()
cenv = ContextualObjectiveFunction(name="LinearContextualSphere",
n_params=2,
n_context_dims=1,
random_state=0)
env2 = SetContext(cenv, context=np.array([0.0]))
env2.init()
assert_equal(env.get_num_inputs(), env2.get_num_inputs())
assert_equal(env.get_num_outputs(), env2.get_num_outputs())
assert_almost_equal(env.get_maximum_feedback(),
env2.get_maximum_feedback())
params = np.zeros(2)
f = np.empty(2)
for i, e in enumerate([env, env2]):
e.reset()
e.set_inputs(params)
e.step_action()
out = np.empty(1)
e.get_outputs(out)
assert_true(e.is_evaluation_done())
f[i] = e.get_feedback()[:]
assert_false(e.is_behavior_learning_done())
assert_almost_equal(f[0], f[1])
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_missing_behavior_search():
ctrl = ContextualController(environment=ContextualObjectiveFunction())
beh = DummyBehavior(initial_params=np.array([0.0]))
beh.init(0, 1)
feedback = ctrl.episode_with(beh)
assert_equal(len(feedback), 1)
assert_less(
feedback[0],
ctrl.environment.get_maximum_feedback(ctrl.environment.context))
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_record_contexts():
contexts = np.linspace(-5, 5, 11)[:, np.newaxis]
class CREPSFixedContextOrder(CREPSOptimizer):
def get_desired_context(self):
return contexts[self.it]
opt = CREPSFixedContextOrder()
ctrl = ContextualController(environment=ContextualObjectiveFunction(),
behavior_search=JustContextualOptimizer(opt),
n_episodes=11,
record_contexts=True)
ctrl.learn()
assert_array_equal(contexts, ctrl.contexts_)
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)
def test_learning_fails_with_missing_behavior_search():
controller = ContextualController(
environment=ContextualObjectiveFunction())
assert_raises_regexp(ValueError, "BehaviorSearch is required",
controller.learn)