Computational Study of Mean-Risk Stochastic Programs

Mean-risk stochastic programs model uncertainty by including risk measures in the objective function. This allows for modeling risk averseness for many problems in science and engineering. This dissertation addresses gaps in the literature on stochastic programs with mean-risk objectives. This includes a need for a computational study of the few available algorithms for this class of problems. The study was aimed at implementing and performing an empirical investigation of decomposition algorithms for stochastic linear programs with absolute semideviation (ASD) and quantile deviation (QDEV) as mean-risk measures. Specifically, the goals of the study were to analyze for specific instances how algorithms perform across different levels of risk, investigate the effect of using ASD and QDEV as risk measures, and understand when it is appropriate to use the risk-averse approach over the risk-neutral one.

We derive two new subgradient based algorithms for the ASD and QDEV models, respectively. These algorithms are based on decomposing the stochastic program stage-wise and using a single (aggregated) cut in the master program to approximate the mean and deviation terms of the mean-risk objective function. We also consider a variant of each of the algorithms from the literature in which the mean-risk objective function is approximated by separate optimality cuts, one for the mean and one for the deviation term. These algorithms are implemented and applied to standard stochastic programming test instances to study their comparative performance. Both the aggregated cut and separate cut algorithms have comparable computational performance for ASD, while the separate cut algorithm outperforms its aggregate counterpart for QDEV. The computational study also reveals several insights on mean-risk stochastic linear programs. For example, the results show that for most standard test instances the risk-neutral approach is still appropriate. We show that this is the case due to the test instances having random variables with uniform marginal distributions. In contrast, when these distributions are changed to be non-uniform, the risk-averse approach is preferred. The results also show that the QDEV mean-risk measure has broader flexibility than ASD in modeling risk.