Machine Learning-Enabled Evolutionary Two-Stage Stochastic Programming

Two-stage Stochastic Programming (2SP) is an effective framework for decision-making and modeling under uncertainty. Some 2SP problems are challenging due to their high dimensionality and nonlinearity. Machine learning can assist in solving 2SP problems by providing data-driven insights and approximations. Evolutionary algorithms are more general and effective methods for handling various 2SP problems by exploiting their structures and features. However, there is still a research gap in combining machine learning and evolutionary algorithms for solving 2SP problems. Therefore, this paper proposes for the first time a Machine Learning-enabled Evolutionary 2SP framework (MLE2SP), which uses machine learning to construct surrogate-assisted evolutionary optimization frameworks for 2SP. It constructs a novel multi-output 2SP surrogate model that considers scenarios and decision variables of both stages for the first time and proposes a data conversion method to handle the high-dimensional decision variables and scenarios. It also proposes a Machine Learning-enabled Differential Evolution Sampling (MLDES) method to update candidate solutions, which extracts knowledge from dominant candidate solutions to guide the evolutionary direction. Moreover, this work provides open sources of linear and nonlinear two-stage stochastic mixed-integer programming problem instances as benchmark test functions. The effectiveness and generality of the proposed algorithm and framework are verified by the test results on the benchmark test functions and a disaster relief logistics problem, which provide a new research direction for designing general and effective two-stage stochastic programming solving frameworks.