When using machine learning techniques for learning a function approximation from given data it is often a difficult task to select the right modeling technique.
In many real-world settings is no preliminary knowledge about the objective function available. Then it might be beneficial if the algorithm could learn all models by itself and select the model that suits best to the problem.
This approach is known as automated model selection. In this work we propose a
generalization of this approach.
It combines the predictions of several into one more accurate ensemble surrogate model. This approach is studied in a fundamental way, by first evaluating minimalistic ensembles of only two surrogate models in detail and then proceeding to ensembles with three and more surrogate models.
The results show to what extent combinations of models can perform better than single surrogate models and provides insights into the scalability and robustness of the approach. The study focuses on multi-modal functions topologies, which are important in surrogate-assisted global optimization.
This paper proposes an experimental methodology for on-line machine learning algorithms, i.e., for algorithms that work on data that are available in a sequential order.
It is demonstrated how established tools from experimental algorithmics (EA) can be applied in the on-line or streaming data setting.
The massive on-line analysis (MOA) framework is used to perform the experiments.
Benefits of a well-defined report structure are discussed.
The application of methods from the EA community to on-line or streaming data is referred to as experimental algorithmics for streaming data (EADS).
Learning board games by self-play has a long tradition in computational intelligence for games. Based on Tesauro’s seminal success with TD-Gammon in 1994, many successful agents use temporal difference learning today. But in order to be successful with temporal difference learning on game tasks, often a careful selection of features and a large number of training games is necessary. Even for board games of moderate complexity like Connect-4, we found in previous work that a very rich initial feature set and several millions of game plays are required. In this work we investigate different approaches of online-adaptable learning rates like Incremental Delta Bar Delta (IDBD) or Temporal Coherence Learning (TCL) whether they have the potential to speed up learning for such a complex task. We propose a new variant of TCL with geometric step size changes. We compare those algorithms with several other state-of-the-art learning rate adaptation algorithms and perform a case study on the sensitivity with respect to their meta parameters. We show that in this set of learning algorithms those with geometric step size changes outperform those other algorithms with constant step size changes. Algorithms with nonlinear output functions are slightly better than linear ones. Algorithms with geometric step size changes learn faster by a factor of 4 as compared to previously published results on the task Connect-4.