Explainability and performance of anticipatory learning classifier systems in non-deterministic environments

Abstract

In the field of Reinforcement Learning, models based on neural networks are highly performing, but explaining their decisions is very challenging. Instead of seeking to open these black boxes to meet the increasing demand for explainability, another approach is to used rule-based machine learning models that are explainable by design, such as the Anticipatory Learning Classifier Systems (ALCS). ALCS are able to develop simultaneously a complete representation of their environment and a decision policy based on this representation to solve their learning tasks. This paper focuses on the ability of ALCS to deal with non-deterministic environments used in reinforcement learning problems, while discussing their explainability. Directions for future research are thus highlighted to improve both the performance and the explainability of the ALCS to meet the needs of critical real-world applications.