Combining neural networks with other prediction techniques

The complexity and the inherent heterogeneity of real world problems are still one of the major challenges in computer science. Due to the necessity of using different data processing technologies general interest in hybrid systems is a fast growing research area. To support the integration of intercommunicating hybrids the paper suggests the use of distributed software architectures. The main advantages of the approach presented in the paper are the encapsulation of different paradigms, the separation of control and domain knowledge and the reduction of the complexity of individual problem solvers. The first section of the paper gives an overview of the state of the art in hybrid processing. A taxonomy of currently known hybrid system approaches is discussed. Because of the special importance of distributed artificial intelligence (DAI) the author examines issues and research directions in this field and concludes with the presentation of DAI as an integrative paradigm. The author then describes a case study. He gives an overview of the domain (economics) and discuss some prediction methods usually used in this area. After introducing some simple economic relationships and a description of how to use neural networks in multivariate prediction the paper shows how connectionist techniques are embedded in a distributed problem solving framework, called PREDICTOR. The last section shows an example in which a hybrid system out performs the homogenous approaches by combining them intelligently. PREDICTOR is a case study of how to design a so called intercommunicating hybrid system