Computer based empirical models for the analysis of conceptual designs

Research is currently in progress to develop a domain independent methodology to generate quantitative information to compare the suitability of a large number of schemes. The basis of the approach adopted is the use of approximate empirical models, generated through unsupervised reflective learning, complemented with traditional analysis techniques. Thus, if appropriate empirical models do not exist an analysis can still be performed. As part of this approach empirical models will be associated with component classes which would then be compared to the scheme description to identify potentially useful models. The final selection being based on minimising the analysis time and keeping the levels of error and of uncertainty to acceptable limits. During reflective learning, which would generally occur when the system is not processing analysis requests, the empirical models, component classes and solution search strategies will be generated and/or optimised to reduce the typical analysis time for problem domains similar to those already encountered. This process would generally involve the structured exploration of the domains of interest using mostly traditional analysis techniques. In this way, a system based on this approach could learn approximate empirical methods of solving a large number of problem classes, without the need for human intervention, starting from a relatively small knowledge base