Noise Handling in Kriging-Based Optimization Algorithms Applied to Sequential Decision Problems in Infrastructure Planning

Abstract

In this paper, we present a comparative study of stochastic Kriging-based optimization algorithms applied to a generic infrastructure planning problem, using direct policy search (DPS) as a heuristic approach. The evaluated problem is defined as a sequential decision problem and involves a generic infrastructure planning scenario under uncertainty, with performance dependent on system and cost model parameters.  The focus is to compare the impact of heterogeneous noise treatment in an optimization framework approaching three algorithms: Minimum Quantile criterion (MQ), stochastic Efficient Global Optimization (sEGO), and Expected Improvement with Reinterpolation (EIR). The MQ algorithm is the only one that does not address information about the noise variance of the stochastic parameter. The results demonstrate that the algorithms presented satisfactory performance, especially those with heterogeneous noise treatment, and show potential for solving complex infrastructure engineering planning problems under uncertainties in a DPS framework.