Summary
International Symposium on Nonlinear Theory and its Applications
2017
Session Number:C0L-B
Session:
Number:C0L-B-1
A Complex-Valued Reinforcement Learning Algorithm Using Complex-Valued Neural Networks
Masaki Mochida, Hidehiro Nakano, Arata Miyauchi,
pp.564-567
Publication Date:2017/12/4
Online ISSN:2188-5079
Summary:
In Complex-valued Reinforcement Learning (CRL), each action-value is represented by a complex value. Then, search history can be naturally included in the argument, while dominance relationships are decided by the amplitude. CRL is effective for the environments with perceptual aliasing. In order to apply larger-problems with the large number of states, this paper introduces the function approximation for the action-value function by using complex-valued neural networks. The simulation results for a benchmark problem are shown.