98 AICOM VolA Nr.2/3 June/September 1991 Task Level Frameworks for Cooperative Expert System Design Markus Stolze Universitaet Zuerich-Irchel, Institutfuer Informatik, Winterthurerstrasse 190, 8057 Zuerich stolze@iJi.unizh.ch It will be shown why expert systems should no longer be designed as autonomous prob- lem-solvers but as cooperative systems. A methodology for designing these coopera- tive expert systems will be presented that combines methods from cognitive engineer- ing with a popular knowledge acquisition approach, namely the the use of task level frameworks. A real world example from the domain of technical troubleshooting will be used to illustrate the four main steps of the methodology. Introduction In most real world applications, the human percep- tion cannot be excluded from an expert system appli- cation. The primary reason for this is that for every sufficiently complex real world task no autonomous problem-solver can be built, because there will al- ways be some relevant factors of the domain that are not included in the knowledge base, and which are therefore not taken into account during problem- solving (Gutknecht et al. 91). Roth&Woods (89) de- scribe a case where this was ignored when building a diagnostic expert system. The result was that the us- ers of this system not only had to diagnose the de- vice they wanted to repair, but they also had to diag- nose the expert system to see whether it was still "on the right track". Therefore the goal of expert systems design cannot be to automate problem-solving. Instead the goal should be to optimize the problem- solving perfor- mance of the joint system of user and expert system. This is the goal of cooperative expert systems de- sign. The basic assumption is that user and expert system differ in their processing capabilities, their general knowledge about problem-solving and their knowledge about the current problem. For example, processing large amounts of data is much better done by a computer system, whereas relating one prob- lem-solving case to another "similar" case.might be easier for a human. For optimal joint problem-solv- ing the tasks must be distributed to the expe,rt system or the user according to their abilities. This should be done in such a way that limitations of one part of the system are neutralized with the help of the other part of the system. Another (sometimes contradictory) goal is to minimize the amount of communication needed between user and expert system. In traditional expert systems, all reasoning is done by the expert system, while the task of the user is to supply data about the problem situation and to exe- cute the actions the expert system proposes (Woods, 86). For most real world applications this task distri- bution is not optimal because the amount of commu- nication needed in this setting is usually very high (since only low level data is communicated) and be- cause some of the problem-solving tasks would match better the abilities of the user. This research was sponsored by TECAN AG Hom- brechtikon, Switzerland.