A fuzzy modular approach to system modeling based on incentive games

System modeling is fundamental to many fields. The complexity and accuracy of the model used and the kind of knowledge it provides are determined by the application. Modular learning systems have many advantages over the alternatives. From an identification perspective, modular models provide knowledge of the functional composition of a system. However, with modular learning systems comes the problem of decomposition: how do we assign sub-problems to modules consistent with their expertise? Horizontal decomposition approaches involve partitioning the operational ranges of the system into different module assignments. The decomposition objective can be achieved by clustering methods that match similar characteristics. Vertical or functional decomposition approaches involve partitioning the problem domain into groupings over the inputs. Such methods are still rare in practice, especially when uncertain, time-varying systems with little a priori information are involved. The modular incentive network we are presenting in this thesis proposes a self-organizing approach to functional modeling. The Modular Incentive Network is organized into a two-level hierarchy with a single leader module at the output and a number of follower local processing modules at the input. Communication among the modules is modeled after a leader-follower game with incentives. All follower modules are assumed to be autonomous learning agents with local objectives. Leader functionality of the supervisor module involves leading the follower learning to desired partitions. The leader is also responsible for decomposing the problem into relevant partitions. Approximate models of learning dynamics are tracked by the leader, based on the observed learning behavior over a selected feature set. Learning metrics are extracted based on a conflict measure in the feature domain and are used to construct incentives to lead the followers. Decomposition of the global target is also performed using the knowledge on follower learning behavior. This thesis includes various simulation illustrations to support the ideas presented. These examples include approximate-learning modeling, incentive-based learning supervision, and self- organizing modular functional modeling of a two-link robot manipulator. Also included in the thesis is a survey of related fields and topics that will enable the reader to follow the concepts introduced