A Parallel Approach for Computing the Expected Value of Gathering Information

It is important for engineering firms to be able to develop forecasts of recommended courses of action based on available information. In particular, engineering firms must be able to assess the benefit of performing information-gathering actions. For example, an automobile manufacturer may use a computer simulation of a hydraulic motor and pump in the design of a new vehicle. The model may contain random variables that can be more accurately determined through expensive information-gathering actions, e.g., physical experiments, surveys, etc. To decide whether to perform these information-gathering actions, the automobile manufacturer must be able to quantify the expected value to the firm of conducting them. However, the cost of computing the expected value of information (through optimization, Monte Carlo sampling, etc.) grows exponentially with the amount of information that is to be gathered and can often exceed the cost of actually gathering the information. Thus, if information decisions are to be addressed algorithmically, there exists a need for novel algorithmic approaches to reduce the computational expense associated with computing the expected value of gathering information. The contribution of this paper is a novel algorithmic approach for approximating the expected value of perfect information (EVPI) for engineering design problems. In this research, we propose to recast the EVPI as a “parametric” problem. The value of recasting the problem is an exponential reduction in the computational complexity. The proposed approach is validated against a Monte Carlo sampling based approach, the traditional approach for solving EVPIs, on an engineering problem. The engineering problem is to compute the expected value of performing physical experiments to gather information about random variables in a computational model the efficiency of an engine and transmission. The results are compared in terms of computational expense and solution accuracy. The results indicate that by recasting the EVPI as a parametric problem the computational expense is reduced drastically while maintaining solution accuracy.