Training neurocontrollers by local and evolutionary search

Training of neural networks by local search such as gradient-based algorithms could be difficult. This calls for the development of alternative training algorithms such as evolutionary search. However, training by evolutionary search often requires long computation time. In this paper, we investigate the possibilities of reducing the time taken by combining the efforts of local search and evolutionary search. There are a number of approaches to combine these search strategies, but not all of them are successful. This paper provides a review of these approaches. Experimental results indicate that while the Baldwinian and the two-phase approaches are inefficient in improving the evolution process for difficult problems, the Lamarckian approach is able to speed up the training process. Moreover, in the case where no local search method is appropriate for learning the desired task directly, this paper demonstrates that allowing the local search to learn another related task can assist the evolutionary search.