Behavioral Partitioning in a Hierarchical Mixture of Experts using K-Best-Experts Algorithm

In recent years methods for combining multiple experts (multi expert systems, MES) have been used to solve different problems of classification and regression. In particular hierarchical mixture of experts (HME) has been widely studied. This paper presents a novel method which divides the problem space into behaviorally portioned subsets using k-best-experts algorithm and then uses the HME structure to assign an expert to each subset. The gates used in the HME structure are support vector machines which are trained to route each problem to the best fitting expert. The method is implemented and tested on the DELVE framework and is compared with other similar methods