Qualitative Models as Indices for Memory-Based Prediction

AT THE SWISS FEDERAL INSTItute of Technology’s Artificial Intelligence Laboratory, we have developed an approach that combines qualitative reasoning and memory-based reasoning, thereby exploiting their strengths and compensating for their weaknesses. We have applied this approach to two processes: coffee roasting and decaffeination. Both applications used large amounts of data collected from plants operated by Nestle in the UK and Spain. Roasting and decaffeination are processes for which existing models can provide only very inaccurate predictions. In both cases, attempts to predict behavior using statistical methods and neural networks have not provided usable predictions. In contrast, the qualitative models used in memory-based reasoning take into account subtleties of the processes that purely statistical criteria are likely to miss. The results are thus significantly better than what conventional methods could produce.