Comparing stochastic design decision belief models: Pointwise versus interval probabilities

Abstract

Decision support systems can either directly support a product designer or support an agent operating within a multi-agent system (MAS). Stochastic based decision support systems require an underlying belief model that encodes domain knowledge. The underlying supporting belief model has traditionally been a probability distribution function (PDF) which uses pointwise probabilities for all possible outcomes. This can present a challenge during the knowledge elicitation process. To overcome this, it is proposed to test the performance of a credal set belief model. Credal sets (sometimes also referred to as p-boxes) use interval probabilities rather than point-wise probabilities and therefore are easier to elicit from domain experts. The PDF and credal set belief models are compared using a design domain MAS which is able to learn, and thereby refine, the belief model based on its experience. The outcome of the experiment illustrates that there is no significant difference between the PDF based and credal set based belief models in the performance of the MAS.