Neural networks for real-time knowledge based applications

The AURA (Advanced Uncertain Reasoning Architecture) project aims to develop a new class of architectures supporting high performance knowledge based systems, particularly in real time applications. The architectures under investigation offer a true integration of symbolic and neural network styles of reasoning in a unified framework. In AURA, the basic functional blocks are neural networks which can perform symbolic reasoning, as well as the pattern processing functions normally associated with neural networks. A further feature of the approach is that a powerful partial match mechanism supports uncertain reasoning. A central theme in the AURA approach concerns the development of techniques which allow inferences expressed as production rules (or predicates), to be implemented as associative mappings. We are also considering specific features required in real time applications, such as anytime functionality. An anytime algorithm is an iterative refinement algorithm that can be interrupted and asked to provide an answer at any time. There are several possible ways in which the AURA architecture can be used in the design of a knowledge based system. One of the simpler possibilities is to use AURA as a component or subsystem of a larger KBS. In that case, the high speed search abilities of AURA would complement many different styles of KBS. However, AURA also forms a basis for a new class of architectures, which perform reasoning tasks using associative mapping as the principal operational concept