An Analysis of ACME : The Analogical Constraint Mapping Engine.

This paper discusses s ome pragmatic issues on the a nalogical constraint m apping engine (ACME), a widely u sed artificial neural network model for analogical matching employing Grossbergs IAC (interactive Activation and Competition) artificial neural network model. Our analysis beings with an investigation into the use of a Hopfield constraint satisfaction n etwork for image reconstruction. This domain providing us with a constraint satisfaction network based on the more usual Hopfield model, but which shares many characteristics with the ACME model. Based on this comparison we demonstrate how the c onvergence time increases exponentially with increasing n etwork size, a factor which h as s erious implications for the practical application o f this model t o all but t rivially small m etaphors. Finally, we briefly present a localist connectionist m odel for solution g eneration in on e domain, with linear scaleability over problem size.