Learning under hardware restrictions in CMOS fuzzy controlers able to extract rules from examples

Abstract

Fuzzy controllers are able to incorporate knowledge expressed in if-then rules. These rules are given by experts or skilful operators. Problems arise when there are not experts or/and rules are not easy to find. Authors' proposal consists in an analog fuzzy controller which accepts structured language as well as input/output data pairs, thus rules can be extracted or tuned from human or software controller operation. Learning from data pairs has to be carried out under hardware restrictions in linearity, range and resolution. In this paper, modelling of building blocks arranged in a neuro-fuzzy architecture is made and issues related to on-chip learning are discussed. Computer simulations show that learning is possible for resolutions up to 6 bits, affordable with the cheapest VLSI technologies.