Neural architecture search and weight adjustment by means of Ant Colony Optimization

Abstract

In recent years, many different areas of research have utilized neural networks (NNs). Many have investigated weights in NNs as well as structural optimization of NNs via neural architecture search. In this paper, we apply weight training during neural architecture search by Ant Colony Optimization(ACO) to two problems from OpenAI Gym and one problem from pybullet-gym controlled by NNs. We also compare the timing of when the weight training is performed, which is before the architecture search, during architecture search and after architecture search. It was found that performing architecture search by ACO and weight training simultaneously is effective for increasing the score of NNs and that by performing weight training before or at the same time as architecture search, the score was increased statistically significantly for all problems compared with fully-connected NN and the score by performing weight training after architecture search was increased statistically significantly for only one problem from pybullet-gym.