Neural networks for safe knowledge based systems

This paper discusses how work at York University is beginning to investigate the use of neural networks for applications where conventional expert systems are inappropriate or fail. A particular emphasis of the approach concerns the problem of dealing with uncertain data. Many applications of expert systems contain uncertainty; often an operator is unable to respond exactly to a request for information, or sensors may be faulty thus attaching some uncertainty to the information they provide. Furthermore, the system running the expert system and the knowledge base may contain hard or soft errors. The inability of expert systems to deal with uncertainty makes them unsafe. The conventional approach is to hand over the task of uncertainty management to a human operator. Unfortunately, this may not be possible in certain circumstances as no human expert may be available or can react in time. This is especially true in safety critical applications. Work at York is looking at the implementation, the application and the theory of uncertain reasoning in neural networks. The eventual aim is to provide tools and methods for applying this new technology in a wide variety of domains