[10p-S422-1~12] 3.10 Optical quantum physics and technologies

[10p-S422-1] [Young Scientist Presentation Award Speech] Model selection based on decision making in photonic reservoir computing

〇Kazutaka Kanno¹, Makoto Naruse², Atsushi Uchida¹ (1.Saitama Univ., 2.NICT)

Keywords:reservoir computing, decision making, semiconductor laser

Reservoir computing is machine learning based on artifcial neural network and its main feature is that only output weights are trained. In time-series prediction based on artifcial neural networks, a trained network may degrade its prediction accuracy when a predicted target is varied due to some changes of environment. To solve this problem, we propose model selection using reinforcement learning in photonic reservoir computing. In this method, two output weights, which are trained using different predicted targets, are used and selected using reinforcement learning. We numerically demonstrate to predict time-series which is temporally switched to different targets.