Metareasoning and meta-level learning in a hybrid knowledge-based architecture

Ahybrid knowledge-based architecture integrates different problem solvers for the same (sub)task through a control unit operating at a meta-level, the metareasoner, which coordinates the use of, and the communication between, the different problem solvers. A problem solver is defined to be an association between a knowledge intensive (sub)task, an inference mechanism and a knowledge domain view operated by the inference mechanism in order to perform the (sub)task. Important issues in a hybrid system are the metareasoning and learning aspects. Metareasoning encompasses the functions performed by the metareasoner, while learning reflects the ability of the system to evolve on the basis of its experiences in problem solving. Learning occurs at different levels, learning at the meta-level and learning at the level of the specific problem solvers. Meta-level learning reflects the ability of the metareasoner to improve the overall performance of the hybrid system by improving the efficiency of meta-level tasks. Meta-level tasks include the initial planning of problem solving strategies and the dynamic adaptation of chosen strategies depending on new events occurring dynamically during problem solving. In this paper we concentrate on metareasoning and meta-level learning in the context of a hybrid architecture. The theoretical arguments presented in the paper are demonstrated in practice through a hybrid knowledge-based prototype system for the domain of breast cancer histopathology.