SYMBOLIC-NEURAL SYSTEMS AND THE USE OF HINTS FOR DEVELOPING COMPLEX-SYSTEMS

Neural network systems can be made to learn faster and generalize better through the addition of knowledge. Two methods are investigated for adding this knowledge: (1) decomposition of networks; and (2) rule-injection hints. Both of these approaches play a role similar to adding rules or defining algorithms in symbolic systems. Analyses explain two important points: (1) what functions which are easy to learn (as well as what functions which make effective hints) are known from an analysis of the effect of learning monotonic functions; (2) a set theory and functional entropy analysis shows for what kinds of systems hints are useful. The approaches have been tested in a variety of settings, and an example application using a lunar lander game is discussed. © 1991 Academic Press Limited.