A MODEL FOR CONCURRENT INFORMATION EXCHANGE BASED UPON THE METHOD OF TEMPORAL DIFFERENCES.

Making the designer aware of downstream impacts of design decisions has been a consistent theme in design and concurrent engineering literature. Mistakes in early design decisions can strongly bias the outcome of the final product although this fact may not become evident until the design is well developed, usually requiring an expensive redesign cycle at that point. We investigate a reinforcement learning approach for providing feedback on downstream effects of early design decisions. The method of temporal differences is discussed as an approach applicable to situations where a sequence of decisions need to be taken before an assessment of the quality of the solution arising from these choices is available. The task in this case is to learn to predict comparative values for intermediate decision alternatives in terms of their ability to lead to a desirable solution. This is presented as a computational analog to a concurrent engineering approach where early design decisions must be evaluated with regard to impact on the final design solution. A simple example in the context of developing specifications for component selection is presented to illustrate the approach.