Integrating Safety Guarantees into the Learning Classifier System XCS

On-line learning mechanisms are frequently employed to implement self-adaptivity in modern systems. With more widespread use in technical systems that interact with their physical environment, e.g. cyber-physical systems, the fulfillment of safety requirements is increasingly gaining attention. We focus on the learning classifier system XCS with its human-interpretable rules and propose an approach to integrate safety guarantees into its rule base. We leverage the interpretability of XCS' rules to internalize the safety-critical knowledge, as opposed to related work, which relies on an external safety monitor. The experimental evaluation shows that such manually injected knowledge not only gives safety guarantees but aids the learning mechanism of XCS. Especially in complex environments where XCS is struggling to find the optimal solution, the use of hand-crafted forbidden classifiers leads to a performance that is up to 41.7 % better than with an external safety monitor.