Neula: a Hybrid Neural-symbolic Expert System Shell

Current expert systems cannot properly handle imprecise and incomplete information. On the other hand, neural networks can perform pattern recognition operations even in noisy environments. Against this background, we have implemented a neural expert system shell NEULA, whose computational mechanism processes imprecisely or incompletely given information by means of approximate probabilistic reasoning. 1 Background Most current artiicial intelligence systems are incapable of handling imprecise and incomplete information. This is a serious drawback, as in most cases it is a totally hopeless task for a programmer to capture all knowledge of the environment. Much research eeort has been expended to improve the robust-ness of artiicial intelligence programs, but very little progress has been made. On the other hand, current neural network research indicates that it is possible to perform pattern recognition operations, such as categorization, even in very noisy environments. We have therefore investigated the interesting question of whether the robustness problem of traditional artiicial intelligence applications can be attacked by neurally inspired techniques for knowledge representation. Inspired by the general idea of hybrid neural-symbolic systems (e.g. 4]), we have introduced in the NEULOG 1 project a knowledge representation scheme for robust storing of hierarchically related concepts. The scheme is based on a neural representation structure , which bears resemblance to a semantic network. However, instead of the traditional artiicial intelligence interpretation of the network as a purely syntactic variant of rst-order predicate calculus, the network performs computations implementing a Bayesian reasoning model, which is capable of inference in the presence of incomplete, imprecise and even inconsistent information .