Learning logic programs with neural networks

Abstract

First-order theory refinement using neural networks is still an open problem. Towards a solution to this problem, we use inductive logic programming techniques to introduce FOCA, a First-Order extension of the Cascade ARTMAP system. To present such a first-order extension of Cascade ARTMAP, we: a) modify the network structure to handle first-order objects; b) define first-order versions of the main functions that guide all Cascade ARTMAP dynamics, the choice and match functions; c) define a first-order version of the propositional learning algorithm to approximate Plotkin's least general generalization. Preliminary results indicate that our initial goal of learning logic programs using neural networks can be achieved.