Controlling the Production of Neuro-symbolic Rules

Abstract

Neurules are a kind of integrated rules integrating neurocomputing and production rules. Neurules can be produced from existing empirical data, through the neurules production algorithm (NPA). In this paper, we present (a) an extension to NPA regarding presentation of neurules, so that they are more natural and more informative, and (b) an experimental comparison of various alternative strategies we can use at some points of NPA targeting at producing as less neurules as possible. Results of (b) show no clear winner for all cases in terms of the gain in number of neurules compared to the computational cost.