Context-based intelligent assistant systems: a discussion based on the analysis of two projects

After the success and rejection of the first expert systems, we are now on the road to designing and developing real intelligent assistant systems, i.e. intelligent systems that use context explicitly. Such systems are called context-based intelligent assistant systems. With the accumulated experience gained from large projects (in our cases two 6-year applications), it is possible to learn from our past failures and the solutions adopted to fill the gaps. First, we developed a successful expert system for a French power company. However, its use in different locations was too strongly constrained by contextual cues. For solving these problems, we concluded that an intelligent system must have a configuration module and a simulation module in addition to the module developed for the main task (e.g. diagnosis or incident management). From a second application for a subway company in Paris (France), we first validate the initial hypothesis about the need of configuration and simulation modules, and then point out the absolute need to make context explicit. This constituted the input for the context-based intelligent assistant system paradigm and its related representation as contextual graphs. In this paper we present the lessons learned from our two applications as well as the main characteristics of context-based intelligent assistant systems with an emphasis on the need to make context explicit. We point out the role of incremental knowledge acquisition, learning and explanation in such context-based systems.