Towards a neurosymbolic approach based on Anticipatory Learning Classifier Systems and Ontologies

This paper aims to achieve a system in which the learning of Anticipatory Learning Classifier Systems (ALCS) is guided by ontologies, allowing us to benefit from the advantages of each approach. ALCS are powerful rule-based models for decision-making in complex, dynamic, and partially observable environments. While their symbolic nature offers strong adaptability, their inductive learning process can lead to invalid or unsafe decisions in domains requiring expert knowledge, constraints, or safety guarantees. To reach our goal, (1) ALCS should be capable of solving increasingly complex learning tasks and (2) a framework enabling us to make ALCS interact with ontologies is needed. We first propose to remove a learning bias present in ALCS. We demonstrate through extensive experiments that comparable performance on standard maze benchmarks is achieved by ALCS without this bias while enabling, for the first time, the solving of the cart-pole problem. We then introduce a novel neurosymbolic approach in which ALCS are coupled with ontologies acting as semantic safeguards. Through symbolic reasoning, the ontology can validate, reject, or suggest alternative actions based on its encoded knowledge, enhancing safety, guiding exploration, and improving interpretability. These contributions pave the way toward knowledge-guided, explainable, and human-in-the-loop learning systems.