Exploring global approximators for multiobjective reservoir control

Efficient multi-purpose reservoir control policies are crucial in the face of frequent and severe floods and droughts, and to balance water allocation across conflicting demands. Evolutionary Multi-Objective Direct Policy Search (EMODPS) is a popular approach to design control policies for multi-purpose reservoir systems. EMODPS, however, relies on experimental choices within the key components of the framework particularly when coupling multi-objective evolutionary optimization with nonlinear approximation networks. This study explores a suite of radial basis functions (RBFs) used to map the system's states to control actions in a flexible manner as time-varying, non-linear relationships. We provide a systematic assessment of different RBF functions to explore their suitability to obtain Pareto efficient control policies. We use the Susquehanna river basin case study in which competing water demands for hydropower, environment, urban water supply, atomic power plant cooling and recreation need to be met. Our findings suggest that the choice of RBF functions have a large impact on the model outcomes and the search behavior of the optimization algorithm.