Representing Arguments as Background Knowledge for Constraining Generalisation

The use of statistical measures to constrain generalisa-tion in learning systems has proved successful in many domains, but can only be applied where large numbers of examples exist. In domains where few training examples are available, other mechanisms for constraining gener-alisation are required. In this paper, we propose a representation of background knowledge based on arguments for and against a hypothesis rather than as statements in logic or probabilistic relations, and show how it can be used to constrain generalisation from single examples (sometimes referred to as`case-based reasoning'). Examples are characterised by the set of arguments for and against a hypothesis of interest, and the resolution of connicting arguments in a current problem is obtained by rstly locating an old example where the same or a similar connict occurred, then secondly generalising the solution in the old example to also apply to the new problem. This allows learning to occur in domains where few training examples exist and background knowledge is available. We provide a description of this method in logical form, and analyse the assumptions under which it is valid, its limitations and possible future extensions.