Evolutionary Multiagent Transfer Learning With Model-Based Opponent Behavior Prediction

This article embarks a study on multiagent transfer learning (TL) for addressing the specific challenges that arise in complex systems where agents have different or even competing objectives. Specifically, beyond essential backbone of state-of-the-art evolutionary TL framework (eTL), this presents novel with prediction (eTL-P) as an upgrade over existing eTL to endow abilities interact their opponents effectively by building candidate models and accordingly predicting behavioral strategies. To reduce complexity models, eTL-P constructs monotone submodular function, which facilitates select Top- ${K}$ from all available based representativeness terms coverage well reward diversity. also integrates social selection mechanisms identify better-performing partners, thus improving performance reducing behavior reusing useful knowledge respect partners’ mind universes. Experiments partner-opponent minefield navigation task (PO-MNT) shown exhibits superiority achieving higher capability efficiency multiple when compared approaches.