A Parallel and Distributed Genetic-Based Learning Classifier System with Application in Human Electroencephalographic Signal Classification

Genetic-based Learning Classifier Systems have been proposed as a competent technology for the classification of medical data sets. What is not known about this class of system is two­ fold. Firstly, how does a Learning Classifier System (LCS) perform when applied to the single-step classification of multiple-channel, noisy, artefact-inclusive human EEG signals acquired from many participants? Secondly and more importantly, is how the learning classifier system performs when incorporated with migration strategies, inspired by multideme, coarse-grained Parallel Genetic Algorithms (PGA) to provide parallel and distributed classifier migration? This research investigates these open questions and concludes, subject to the considerations herein, that these technological approaches can provide competitive classification performance for such applications. We performed a preliminary examination and implementation of a parallel genetic algorithm and hybrid local search PGA using experimental methods. The parallelisation and incorporation of classical local search methods into a genetic algorithm are well known methods for increasing performance and we examine this. Furthermore, inspired by the significant improvements in convergence velocity and solution quality provided by the multideme, coarse-grained Parallel Genetic Algorithm, we incorporate the method into a learning classifier system with the aim of providing parallel and distributed classifier migration. As a result, a unique learning classifier system (pXCS) is proposed that improves classification accuracy, achieves increased learning rates and significantly reduces the classifier population during learning. It is compared to the extended learning Classifier System (XCS) and several state of the art non-evolutionary classifiers in the single-step classification of noisy, artefactinclusive human EEG signals, derived from mental task experiments conducted using ten human participants. We also conclude that establishing an appropriate migration strategy is an important cause of pXCS learning and classification performance. However, an inappropriate migration rate, frequency or selectiomreplacement scheme can reduce performance and we document the factors associated with this. Furthermore, we conclude that both EEG segment size and representation both have a significant influence on classification performance. In effect, determining an appropriate representation of the raw EEG signal is tantamount to the classification method itself.