Guiding Complex Design Optimisation Using Bayesian Networks
John Leaney, Artem Parakhine · Sciyo eBooks · 2010
The design optimisation guidance methodology aims to aid the designer in directing the overall system optimisation process. One of the major difficulties of providing such guidance is the nature by which this optimisation process is advanced. Specifically, the designer is essentially incapable of affecting the qualities directly. Instead, he or she is forced to consider a set of choices targeting the specific features of the design contributing towards achievement of desirable system qualities. As a result, since a single choice could affect multiple qualities, this introduces a requirement for guidance to provide the designer with understanding of the causal relationships existing in the system. Achieving this involves the study of assumptions held by the designer and other stakeholders, the relevance of existing knowledge and the accuracy of possible predictions. The fusion of simulation modelling and the BBNs can serve as tool of such study as its aim is to provide a tangible link between the way in which the system is structured and its observed levels of quality. Additionally, by combining the hybrid simulation model with BBN discovery algorithm we managed to obtain a much more repeatable output that is validated against encoded assumptions and is less prone to human error. However, the method’s success relies greatly on validity of the model and clarity of the BBN representation. To this end we have found that the simulation model should be built in an incremental manner using a variety of information sources and explicit encoding of assumptions help by the participants. Consequently, the extracted BBN plays a dual role both as a guidance tool and a model verification tool as the conditional probabilities it displays can quickly highlight inconsistencies within the model. The results presented herein warrant further investigation along four major axis: • further research is needed to help the designer with choice of quality factors and criteria that contribute to the nodes of the CQM; • a taxonomy of simulation primitives needs to be developed to aid the designer with construction of hybrid simulation models; • additional research is needed to examine how various BBN discovery algorithms perform on the types of simulation output produced by models of systems from different domains; • studies should be conducted into the various stochastic methods of optimisation such as Cross-Entropy (Caserata & Nodar, 2005) that could be implemented based on the outcomes of BBN use for qualitative applied over a succession of system development cycles. Finally, the development of this approach to guidance should be used to construct a fully fledged decision support and optimisation framework described in Section 3.