EditEr: a combination of IEA and CEA
Qiangfu Zhao · 2002
This paper studies the evolutionary learning of neural networks that can be decomposed into many homogeneous modules, and proposes a new algorithm by combining the individual evolutionary algorithm (IEA) and the co-evolutionary algorithm (CEA). The proposed algorithm has two parts. The first part, a modified version of the IEA, consists of four basic operations: evaluation, deletion, insertion and training. This part is to construct the system using as less modules as possible. The second part is CEA, and the purpose of this part is to evaluate and reproduce good candidate modules for constructing the system. The algorithm is called EditEr in this paper. In the EditEr, an individual is assigned to each module, and the fitness of an individual is defined according to its contribution to the system; a population is assigned to each class of individuals, and many individuals are to be found from each population. Some experimental results are provided to show the efficiency of the EditEr.