EE Seminar: Multi-objective Topology and Weight Evolution of Neuro-controllers and its Application for Robot Navigation
Speaker: Omer Abramovich
M.Sc. student under the supervision of Prof. Emilia Friedman & Dr. Amiram Moshaiov
Wednesday, March 22nd 2017 at 15:30
Room 011, Kitot Bldg., Faculty of Engineering
Multi-objective Topology and Weight Evolution of Neuro-controllers and its Application for Robot Navigation
Abstract
In the last decades, there has been a substantial increase of interest in the field of evolutionary neural networks with an emphasis on evolutionary robotics. Neuroevolution combines the adaptation power of neural networks with the advantages of evolutionary search to find networks capable of solving highly complex tasks.
Evolving neuro-controllers for robot applications might be a hard challenge for standard evolutionary search techniques. Neuroevolution aims to provide some tailored algorithms to such problems. One example is NEAT, which is an evolutionary algorithm to find not only the optimal connection weights, but also the optimal topology of a network. However, such TWEANN (Topology and Weight Evolution of Artificial Neural-Networks) algorithms are commonly restricted to single objective evolution, and their adjustment for multi-objective problems is in its infancy stage.
In this thesis, we propose a novel fully automated approach to evolve topologies and weights of neuro-controllers for multi-objective problems and in particular for navigation application. The suggested approach dismisses any prerequisites of expert knowledge regarding the network topology and enables to achieve better performances than expert designed neuro-controllers. A key component of this study is the proposed modification to NEAT, which results in an innovative generic multi-objective TWEANN algorithm. The suggested algorithm is proven to outperform state-of-the-art multi-objective TWEANNs in the tested cases of this study.
