An Unofficial FastLAS Tutorial: A Programmer’s Guide

Overview of FastLAS

FastLAS is a tool for Inductive Logic Programming (ILP) that learns logic programs from examples. It operates by taking four inputs: background knowledge (fixed facts and rules), a language bias (mode declarations defining the allowed vocabulary), a set of examples (positive and negative scenarios), and a scoring function (a bias defining what makes a hypothesis 'better'). The system is designed to scale to tens of thousands of examples by avoiding the exhaustive enumeration of the entire hypothesis space, instead focusing on candidate rules that are relevant to the provided data.

The Learning Process

FastLAS computes a hypothesis that covers all 'hard' examples while minimizing a cost function. This cost is derived from the user-defined #bias, which can penalize rule length or specific predicates. The system distinguishes between two learning algorithms: --opl (Observational Predicate Learning), which assumes the target predicate is present in the examples, and --nopl (Non-Observational Predicate Learning), which is more general but computationally more expensive. The system is built on top of the Clingo ASP solver, meaning it leverages stable model semantics to evaluate whether a learned hypothesis explains the given examples.

Practical Implementation

Writing a FastLAS program involves defining the structure of the rules through #modeh (head) and #modeb (body) declarations. These declarations act as templates, and the #maxv directive limits the number of variables allowed in a single rule. Because FastLAS is not a full implementation of the Clingo language, users must be careful to avoid unsupported syntax like conditional literals. Instead, complex logic must be refactored into simpler, non-recursive rules using negation as failure.

FastLAS provides a bridge between raw data and symbolic logic, allowing researchers to automate the discovery of rules in domains like access control or event recognition. By providing a 'programmer's guide' approach, this paper enables users to move beyond simple examples and effectively tune the system's performance for large-scale, real-world tasks where traditional ILP systems might struggle with memory or time constraints.