Rapid Prototyping of ILP Systems Using Explicit Bias

We describe a system that allows one to rapidly implement a variety of biases for inductive logic programming; this is an especially suitable area for rapid prototyping, as a number of diierent restrictions on logic programs have been proposed. Our prototyping system is an extension of the Grendel learning system, which accepts two inputs: a set of examples of the concept to be learned, and an explicit description of the desired bias, expressed in a \bias representation language". Rapid prototyping can be accomplished using Grendel by varying the bias description and keeping the rest of the learning system xed. To allow rapid prototyping for inductive logic programming biases, we have extended Grendel's bias representation language with a powerful macro facility called \lazy macros" that allows succinct descriptions of a number of useful biases. In particular, we present descriptions of the following biases, and compare them experimentally: constant-depth determinate clauses, clauses with a constant number of \free" variables, clauses found via \relational pathhnding", and a previously unexplored bias toward clauses with constant \locality". The last bias is suggested by recent formal results.