The first presentation in this session was given by Peter Bosman from CWI, Amsterdam. He presented a joint work with Dirk Thierens from Utrecht University. The title of the paper is: “More Concise and Robust Linkage Learning by Filtering and Combining Linkage Hierarchies”. The research extends previous approaches on linkage learning in which models are used to group different variables in thechromosome together. Dr. Bosman first explained several linkage learning methods and associated representations. Then he showed how they can be combined in one system, hopefully gaining the advantages from the subsystems. Many experimental results on several challenging problems showed that the resulting combined system can be more robust than single systems.
The second presentation was given by Madalina Drugan from the Vrije Universiteit Brussel, who presented the paper “Designing Multi-objective Multi-armed Bandit Algorithms: a Study”, which is joint work with Ann Nowe. Dr. Drugan outlined the need for multi-objective decision making in some applications and presented a number of different algorithms that can be used to find the set of Pareto optimal arms. Furthermore, she showed a very thorough theoretical analysis of the different approaches. At the end some experimental results were given showing that the Pareto-based multi-armed bandit method clearly outperforms the other approaches.
The last presentation in the session was given by Saba Yahyaa who presented joint work with Bernard Manderick, both from the Vrije Universiteit Brussel. The presentation was entitled “Knowledge Gradient Exploration in Online Kernel-Based LSPI. Ths speaker first covered the field of reinforcement learning, and then continued with the use of least-squares policy iteration (LSPI). After this the use of kernels in the LSPI algorithm was described and finally this algorithm was combined with a novel exploration method called Knowledge Gradient. The method was used on a maze problem and the results showed the effectiveness of the approach.
