A schema-based model of adaptive problem-solving
Roy M. Turner, Janet L. Kolodner · SMARTech Repository (Georgia Institute of Technology) · 1989
Problem solving research in artificial intelligence (AI) has traditionally looked at problems in closed domains in which the reasoner knows enough to predict all changes to the problem-solving situation ahead of time. Unfortunately, in most interesting domains, a reasoner cannot assume it has complete knowledge. Instead, it must be able to adapt its behavior to fit the situation it is in and to cope with changes to the situation as they occur. We have developed a model of adaptive reasoning called schema-based reasoning (SBR). SBR grew out of two disparate trends in recent AI research: case-based reasoning and reactive planning. Our approach solves problems using schematic knowledge (schemas) representing generalized problem-solving sessions and portions of sessions. Procedural schemas (p-schemas) represent generalized action sequences and play a role analogous to plans in other systems: achieving goals. Contextual schemas (c-schemas), representing generalized cases of problem solving, provide knowledge about the problem-solving context that allows appropriate responses to changes and allows the reasoner to focus its attention on appropriate goals to pursue. Strategic schemas (s-schemas), representing generalizations of the abstract features of cases, help control reactivity and help the reasoner focus its attention. Schema-based reasoning proceeds largely by recognition. When the reasoner recognizes the current situation as an instance of a kind of context it knows about, it uses the corresponding c-schema to control its behavior. When it recognizes a goal as one it knows how to achieve, it finds an appropriate p-schema. And when it recognizes the abstract features of the problem-solving situation as implying that it should use a particular strategy, the corresponding s-schema is found and used. SBR is a reactive, opportunistic reasoning method. P-schemas are interruptible, and c-schemas and s-schemas allow the automatic selection of appropriate responses and of goals to pursue. Though not a primary focus of this research, SBR also has provisions for using past experience to guide problem solving. Indeed, schemas can be viewed as generalized cases of past problem solving. Our model of adaptive reasoning is implemented in the MEDIC program, a schema-based medical diagnostic consultant whose domain is pulmonology.