Incentive games for neuro-fuzzy control

Abstract

Introduces a two-level modular neuro-fuzzy network based on incentive games where the modules are organized as autonomous local optimizers in a leader-follower game hierarchy. Incentive-reaction pairs are used as a measure for the capacity and responsiveness assessment of each follower module. Learning within the follower modules is performed in a traditional error-based manner (e.g. backpropagation). The allocation of targets and incentives to each follower module, on the other hand, is independent of connection weights; incentive games are used for that purpose. Two important advantages of the new architecture are its physically significant follower module outputs and the context-based enhancement it makes to backpropagation.