Facts and Fallacies in Using Genetic Algorithms for Learning Clauses in First-Order Logic

Over the last few years, a few approaches have been proposed aiming to combine genetic and evolutionary computation (GECCO) with inductive logic programming (ILP). The underlying rationale is that evolutionary algorithms, such as genetic algorithms, might mitigate the combinatorial explosions generated by the inductive learning of rich representations, such as those used in first-order logic. Particularly, the binary representation approach presented by Tamaddoni-Nezhad and Muggleton has attracted the attention of both the GECCO and ILP communities in recent years. Unfortunately, a series of systematic and fundamental theoretical errors renders their framework moot. This paper critically examines the fallacious claims in the mentioned approach. It is shown that, far from restoring completeness to the learner progol’s search of the subsumption lattice, the binary representation approach is both overwhelmingly unsound and severely incomplete.