ASAFES2: a novel, neuro-fuzzy architecture for fuzzy computing, based on functional reasoning

Abstract

The functional reasoning or Takagi-Sugeno fuzzy reasoning method is a very promising approach to the task of multi-variable, non-linear function approximation, and the design of a fuzzy model and fuzzy controller. The proposed architecture, ASAFES2, is a neuro-fuzzy function approximator, and is the first attempt to combine this reasoning method with stochastic reinforcement learning. It can “learn” a function to a high degree of accuracy, at a very great speed, and furthermore, expresses it in a set of linear relations. The main ideas are the fuzzy partitioning of the input space into fuzzy subspaces (each corresponding to a possible fuzzy rule), and the use of a separate neural unit for every fuzzy subspace, in order to calculate the optimum consequence parameters. Simulation results prove its superiority over back-propagation in simple non-linear function approximation tasks, and over all previous approaches to the Box and Jenkins gas furnace modelling. Comparisons are also made with Generalized Radial Basis Function Networks. Finally, ASAFES2, as a building block in a learning fuzzy logic controller, is applied to the well-known cart-pole stabilization problem, with satisfactory results. A new, flexible membership function is also presented.