Building competent reflective systems

A competent system need not be perfect. It should not try to solve unsolvable problems. Rather it should simplify and decompose overcomplex problems, complement missing information, weight the quality of a solution against the cost of achieving it, approximate solutions if time is sparse, etc. To make a knowledge-based system more competent, we propose to enhance it by suitable competence specialists, each being devoted to a special type of competence assessment and improvement. The resulting system is a reflective system since it reasons about and modifies its own problem-solving behavior. We developed a framework and the language MODEL-K to describe such systems at the knowledge level and implement them afterwards in a structure preserving way. We applied this framework to improve a knowledge-based system for assignment problems by ten competence specialists. 1 What is competence? Knowledge-based systems are used to solve difficult problems. But they often fail abruptly because they had no idea of their own abilities. Competence does not mean that one knows everything or that one makes no errors. Competent systems however know what they do not know and they recognize errors and deadlocks in their approach as fast as possible [22]. They can • estimate whether a problem is solvable • modify a problem to make it solvable • assess the effect of missing or uncertain information • simplify problems before solution finding • estimate the cost of solution finding • compare different strategies • recognize deadlocks and search for ways out • classify problems and compare them with each other • request support from outside • handle time restrictions. This article originates from a research project which was partly funded by the ESPRIT Basic Research Programme of the Commission of the European Communities under no. 3178. The partners of this project were the University of Amsterdam (NL), the GMD (D), the Dutch Energy Research Foundation ECN (NL) and BSR-Consulting (D). The above shows that, in addition to pure problem solving, there are mechanisms which are not used for actual solution finding, but to monitor one’s own approach from a higher level and which intervene if required. A competent system will reject unsolvable problems, for example, or it will reduce them to make them solvable. The integration of meta knowledge into knowledge-based systems has already been discussed very early in AI [18] [4] [5]. Typical meta activities considered were strategic control [8] and meta planing [23] [17], recognition and the repair of deadlocks in problem-solving [15] [18] [9], explanation of knowledge bases [4], and acting under time restrictions [16]. In logic, too, meta levels have been studied (see [19] for an excellent overview), in particular wrt. self-referential phrases [14].