Modular neural networks with applications to pattern profiling problems

We study the feasibility and the performance of modular design concept as applied to pattern profiling problems using artificial neural network. By decomposing the given pattern profiling problem into smaller modules, it is shown that comparable performance can be achieved with improvement on computation and design complexity. A survey of typical modular neural networks shows that large-scale nonlinear problems can alleviate its dimensionality curse with modular technique. Overview of modular neural networks based on how the problem is modularized through various decomposition and subsequent aggregation is given. A pattern recognition problem for aircraft trajectory prediction using NeuroFuzzy learning with a two stage modular learning design is presented. Decoupled data are used to train respective neural network modules. A genetic algorithm is used to aggregate all the learned modules so that it is ready for online pattern recognition purpose. As compared with the non-modular approach, the modular approach offers comparable prediction performance with significantly lower overall computation time. This study validates that modular design is a promising solution for large-scale soft computing problems.