A proposal of 2D artificial cellular neural network including nonvolatile unit
Daisuke Hiratsu, Tomoharu Nagao · Systems and Computers in Japan · 2002
Abstract We have investigated the evolutional optimization of the Artificial Cellular Neural Network (ACNN). ACNN is a kind of neural network model. Neural units in ACNN are placed on a lattice grid, connections in ACNN are allowed only between neighborhood units, and connection weights and threshold parameters can be set independently of other units. In ACNN, connection between far units does not exist. Thus, ACNN architecture is suitable for building as hardware. We have suggested an optimization method of ACNN using a genetic algorithm (GA). In this paper, we introduce the “nonvolatile” attribute into partial units of a 2D ACNN in order to get the ACNN to output suitable data in a certain situation that has a causal relationship with past input. As its exercise, we used Sutton's maze problem, which is well known in the field of reinforcement learning. We prepared a simulation environment in which a 2D ACNN makes decisions of action in Sutton's maze to demonstrate that a 2D ACNN which includes nonvolatile units has better signal processing ability than the usual 2D ACNN. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(7): 40–49, 2002; Published online in Wiley InterScience ( www.interscience.wiley.com ). DOI 10.1002/scj.1140