One framework, two systems: flexible abductive methods in the problem-space paradigm applied to antibody identification and biopsy interpretation

Abstract

Our goal is to build flexible knowledge-based systems which can use a variety of problem-solving methods and additional task knowledge, without altering the method or task representation. For this purpose, we use a problem-space architecture which allows opportunistic adaptation of problem-solving methods based on the particular goal, situation and knowledge available. Within this framework, we have developed an opportunistic problem-solving method for flexible abductive problem solving. The basic method was specified in terms of potential subgoals and preferences regarding the order of subgoals. This technique avoids specification of any unnecessary procedural commitments, making the resulting method very general and robust. We then developed two systems using this basic abductive method with additional domain-dependent search-control and task knowledge. The additional knowledge alters/overrides some of the minimal knowledge provided by the basic method. Behavior of these systems can change quite dramatically depending on the added increments of knowledge. Knowledge available at runtime shapes the method and hence the behavior. Moreover, even in the absence of strong domain knowledge, due to the wide coverage of the basic abductive method and the general architecture it is based on, the systems are always able to perform some problem solving--they are not brittle. The specific systems implemented are RedSoar and LiverSoar in the domain of red blood cell antibody identification and liver biopsy interpretation, respectively. We discuss the designs, implementations and evaluations of these systems, emphasizing the role the flexible abductive problem-solving method plays.