Seminar - Cooperative Convergent Evolution with Multiple Learning Classifier Systems

ECS PhD Proposal

Speaker: Yi Liu
Time: Friday 24th August 2018 at 12:00 PM - 01:00 PM
Location: Cotton Club, Cotton 350

Add to Calendar Add to your calendar

Abstract

Learning classifier systems (LCSs) have been successfully adapted to many domains with the claim of human-readable rules. However, due to the inherent rich characteristic of the employed representation, it is possible to represent the underlying patterns in multiple (polymorphic) ways, which obscures the most informative patterns. Three hypotheses drive this project's objectives. Firstly, in LCSs' trained agents, removing the inherent diversity and polymorphism through convergent evolution will lead to more compact solution. This will increase the efficiency and in certain cases will result in a more effective classification. Secondly, in domains that are prone to over-general rule formation, LCSs have problems in directly removing over-general rules. Adapting the mutation operators to specialize over-general rules, instead of removing them, will likewise improve classification. Thirdly, it is hypothesized that decision boundaries are transferable from small to more complex environments. Instead of transferring the identified patterns of features in solution space, methods will be developed to transfer patterns of features' importance in sample space. This will counteract the increasing search space of solutions in rich representations as the domain scales. In preliminary work, two rule reduction algorithms and an ensemble learning based LCS have been developed. Afterwards, adapting ensemble learning to LCSs, adapting LCSs to over-lapping domains, adapting transfer learning to LCSs in the bottom-up manner of the sample space scaling, and adapting deductive learning to LCSs are proposed in sequence in order to improve LCSs' power in both adaptability and interpretability.

Go backGo back to the seminar list