Setsuo Tsuruta, Masaki Honma, Hideaki Ishida and Akio Nakano
A considerably complex expert system called elevator maintenance engineer scheduling AI system was developed and has been practically used for more than seven years. However, its knowledge refinement and resultant system validation has been constantly required for its survival as a practically usable system. Their cost is quite expensive compared with the validation/refinement cost needed for ordinary software systems or simple expert systems. In this paper, the history of above refinement and validation is described and the problems of the AI system validation and its knowledge refinement are analyzed. Based on this experience, some ideas to overcome these problems are proposed for efficiently building the right or practically usable AI system.