Connection Sparsification and Orbit Stabilization of Dynamic Binary Neural Networks based on Multiobjective Evolutionary Algorithms
Tomoyuki Togawa, Toshimichi Saito · 2020
0 A dynamic binary neural network is characterized by ternary connection parameters and can generate various bi-nary periodic orbits. This paper studies a two-objective problem for sparsity of the connection parameters and stability of the binary periodic orbits. In order to optimize the two-objective problem, we present a simple algorithm (ALG/M) based on the multiobjective evolutionary algorithm based on decomposition. The ALG/M decomposes the two-objective problem into multiple subproblems and can optimize the problem effectively. Performing elementary numerical experiments for typical examples of binary periodic orbits, it is confirmed that the ALG/M realizes both appropriate connection sparsity and strong orbit stability. It is also confirmed that the ALG/M outperforms another algorithm based on the regularization algorithms such as the Lasso.