Comparison of symbolic and connectionist approaches to local experts integration

Monostrategy classification systems are very limited in the type of knowledge they can use for decision making. On the other hand, potentially better results are achievable using multistrategy systems that integrate two or more types of knowledge representation and/or multiple inference underlying a decision process. In this paper a decision tree and a neural network technique for competitive integration of heterogeneous local experts are proposed. The local experts are either symbolic rule-based classifiers or neural network based monostrategy learning systems. The integration is simple, as it involves no modification of existing symbolic components. The proposed competitive integration systems are tested versus a previously used cooperative neural network based approach. The experimental results on a small financial advising problem indicate significant performance improvements when using the neural network based competitive integration approach as compared to the results obtained from either the individual classifiers, a decision tree based symbolic integration or a cooperative neural network integration method. The best competitive neural network results are achieved by incorporating prior knowledge and a dynamic neural network local expert into the integrated system