Ejemplo n.º 1
0
    def rollout_with_model_performance(self, history=[]):
        value_to_choose = []
        list_configuration_to_choose = []
        list_rollout = []
        st_time = time.time()
        try:
            configs = self.config_space.sample_partial_configuration(
                history, 500)
            list_configuration_to_choose = [
                np.nan_to_num(c.get_array()) for c in configs
            ]
        except Exception as e:
            pass
        mu, sigma = self.score_model.get_mu_sigma_from_rf(
            np.array(list_configuration_to_choose), self.score_model.model)
        ei_values = expected_improvement(mu, sigma,
                                         self.bestconfig["validation_score"])
        mu_time, sigma_time = self.score_model.get_mu_sigma_from_rf(
            np.array(list_configuration_to_choose),
            self.score_model.model_of_time)
        ei_values = [
            ei if mu_t < (self.cpu_time_in_s) else -1000
            for ei, mu_t in zip(ei_values, mu_time)
        ]

        id_max = np.argmax(ei_values)
        return configs[id_max]
Ejemplo n.º 2
0
    def rollout_in_expert_neighborhood(self, history=[]):
        try:
            expert, _ = self.experts[tuple(history)]
        except:
            expert = self.config_space.sample_partial_configuration_with_default(
                history)

        try:
            #configs = [c_f for c_f in
            #    [get_one_exchange_neighbourhood_with_history(c_i, self.seed, history) for c_i in
            #        [c for c in get_one_exchange_neighbourhood_with_history(expert, self.seed, history)]
            #]]
            configs = []
            # print("[Rollout] Neighborhood on", expert.get_dictionary())
            for c in get_one_exchange_neighbourhood_with_history(
                    expert, self.seed, history):
                tmp_ = list(
                    get_one_exchange_neighbourhood_with_history(
                        c, self.seed, history))
                configs.extend(tmp_)
            configs.extend(
                self.config_space.sample_partial_configuration(history, 1000))

            list_configuration_to_choose = np.nan_to_num(
                np.array([c.get_array() for c in configs]))

        except Exception as e:
            raise (e)

        mu_loc, sigma_loc = self.score_model.get_mu_sigma_from_rf(
            np.array(list_configuration_to_choose), "local")
        ei_values = expected_improvement(mu_loc, sigma_loc,
                                         self.bestconfig["validation_score"])
        #mu_time, sigma_time = self.score_model.get_mu_sigma_from_rf(np.array(list_configuration_to_choose), "time")

        #if self.is_metalearning:
        #    beta = self.get_beta()
        #    mu_gen, sigma_gen = self.score_model.get_mu_sigma_from_rf(np.array(list_configuration_to_choose), "general")

        #ei_values = np.array([mu_gener * beta + (1 - beta) * mu_local for mu_gener, mu_local in zip(mu, mu_loc)])

        id_max = (-ei_values).argsort()[:3]
        return [configs[np.random.choice(id_max)] for _ in range(3)]