EE Seminar: Multi-objective Co-Evolution of CPN Controllers for a One-on-one Robotic Soccer Game
~~Speaker: Naftali Kopilevich,
M.Sc. student under the supervision of Prof. Emilia Fridman and Dr. Amiram Moshaiov
Monday, July 27, 2015 at 15:00
Room 011, Kitot Bldg., Faculty of Engineering
Multi-objective Co-Evolution of CPN Controllers for a One-on-one Robotic Soccer Game
Abstract
Counter Propagation Neuro-controllers (CPNs) are trained by co-evolution for the purpose of a robotic soccer-like game. In contrast to the commonly used Feed-Forward Neuro-controllers (FFNs), CPNs are characterized by a Kohonen layer, where the inputs are self-organized. This is followed by a Grossberg layer, in which a mapping of the input classes to the control outputs occurs.
The Kohonen layer is trained ahead of the evolutionary process by either self-organizing or alternatively by k-means clustering. Next, the Grossberg layer undergoes a co-evolutionary process. Numerical simulations are carried out using two known co-evolutionary schemes. The co-evolution results are compared and discussed.
This work shows, within the limitations of the considered case, that CPN-based neuro-controllers are comparable with FFN-based ones. It also emphasizes that a two-phase co-evolutionary scheme, with a multi-objective 1st phase, has a considerable run-time advantage over a single-objective one-phase scheme. Finally, future work is suggested.