GECCO2021 - tut138 - Advanced Tutorials - Advanced Learning Classifier Systems

Advanced Learning Classifier Systems (tut138, Advanced Tutorials) Will Neil Browne This tutorial presents how advanced Learning Classifier Systems embed Evolution with Machine Learning to produce transparent, explainable and flexible AI systems. Evolutionary Machine Learning is continuing to grow in popularity due to its combination of efficient genetic search with effective machine learning. Although this is led by Evolutionary Deep Learning and Neuroevolution, where evolution seeks to find the frameworks/parameterisation for connectionist approaches, there are alternatives. Learning Classifier Systems adopt more of a symbolic approach through evolving heuristics, which helps advance explainable AI. Furthermore, LCS are over 40 years old as a concept, where just as artificial neural network approaches have advanced in this time, so have LCS. The reasons why many concepts associate with LCS are now considered superseded are presented, where instead modern reinforcement learning-based updates are considered. Similarly, instead of old/restricted representations being discussed, such as the ternary alphabet, the most recent advances in knowledge perception are described for adaptation to real-world domains. How LCS can be adapted to problems ranging from bioinformatics to robotic navigation will be explained. How to adapt LCS to future academic research directions will be considered, for example continual learning. How hybrid methods with other EC techniques, such as Genetic Programming, for feature selection and feature construction will be discussed. Inspiration from natural cognitive learning will be used to motivate directions for further LCS enhancements. GECCO 2021 The Genetic and Evolutionary Computation Conference July 10-14, 2021 — Lille, France (online) https://gecco-2021.sigevo.org