To appear: International Journal of Human-Computer Studies, 1998, 49, pages 867-893 Towards Situated Knowledge Acquisition Tim Menzies NASA/WVU Software Research Lab, 100 University Drive, Fairmont, WV USA tim@menzies.com Abstract Situated cognition is not a mere philosophical concern: it has pragmatic implications for current practice in knowledge ac- quisition. Tools must move from being design-focused to be- ing maintenance-focused. Reuse-based approaches (e.g. us- ing problem solving methods) will fail unless the reused de- scriptions can be extensively modified to suit the new situ- ation. Knowledge engineers must model not only descrip- tions of expert knowledge, but also the environment in which a knowledge base will perform. Descriptions of knowledge must be constantly re-evaluated. This re-evaluation process has implications for assessing representations 1 . Introduction Consider a knowledge base which is conceived, built, and used. How can we improve our skills for developing such knowledge bases? If most of the changes to that knowledge base occur be- fore its usage, then we would look to optimising the de- sign process: i.e. conception to construction. Standard knowledge acquisition practice (e.g. KADS (Wielinga, Schreiber, & Breuker 1992; Breuker & de Velde (eds) 1994)) is very focused on optimising design. If most of the changes to that knowledge base occur once it is being used, and we should look to optimis- ing the maintenance process: usage to reconception to reconstruction. Very few knowledge acquisition tools are maintenance-focused (exception: ripple-down-rules (Compton & Jansen 1990)). The claim of this article is that is if we accept situated cog- nition, then knowledge acquisition must move away from current approaches which seek to optimise design. That is, situated cognition challenges established current design- focused knowledge acquisition practice. It could be argued that it is a false dichotomy to separate an emphasis on design from an emphasis on maintenance. Isn’t good design the best means of simplifying mainte- nance? Perhaps not. As we shall see below, certain suc- cessful maintenance-focused tools can take a very minimal approach to initial design. 1 This work was performed while the author was located at the Artificial Intelligence Department, School of Computer Science and Engineering, University of NSW, Australia, 2052 It could also be argued that a philosophical perspective on human reasoning has little relevance for tool builders such as pragmatic knowledge engineers. Situated cognition seems a notion of science, ‘what is the basis of natural human behav- ior’, while knowledge acquisition is a notion of engineer- ing, ‘how can we build cost-effective systems’. However, a knowledge base is an embodiment of some scientific model of the world. We build models to explicate and share our understanding of a domain. Idiosyncrasies in our conceptu- alisation process should be managed in our modeling tools. This article focuses on the idiosyncrasy of changing our minds. People, even experts, change their mind frequently. Due to situated cognition, a description of some treasured belief from today may be different than yesterday’s descrip- tion of that belief. Will this difference significantly disturb the modeling process? Later in this article, we will apply these two tests to check if situated cognition is a pragmatic concern for tool builders: Test 1: Design-focused knowledge acquisition techniques assume that abstracted portions of old knowledge bases (ontologies (Gruber 1993) or problem solving methods (Breuker & de Velde (eds) 1994; Chandrasekaran, John- son, & Smith 1992)) can be reused for new applications. Situated cognition is a major concern for current practice if knowledge base changes preclude the successful appli- cation of these abstracted portions of old knowledge bases to new situations. Test 2: Situated cognition claims that we should expect descriptions of knowledge (e.g. an ascii knowledge base) to undergo major change during its lifetime. Situated cog- nition is a pragmatic concern if knowledge base change is a major issue for real-world knowledge bases. Vera and Simon offer other arguments against chang- ing current practice due to situated cognition (Vera & Si- mon 1993b; 1993a; 1993c). Firstly, they argue that the physical symbol system hypothesis (Newell & Simon 1972) has been a fruitful paradigm which can reproduce many known behaviours of experts. This is a compelling argu- ment. Why should we change current practice when current practice has produced so many successful expert systems (e.g., PROSPECTOR (Campbell et al. 1982; Duda, Hart, & Reboh 1985), XCON (Bachant & McDermott 1984), VT (Marcus, Stout, & McDermott 1987), PIGE (Menzies et 1