Balancing Exploration, Uncertainty and Computational Time in Many-Objective Reservoir Policy Optimization

As we confront the challenges of managing river basin systems with a higher number of reservoirs and increasingly uncertain tradeoffs impacting their operations, we will need more computationally efficient optimization strategies. Future evolutionary multi-objective direct policy search (EMODPS) methods must scalably address the rapid growth in the computational demands associated with simulating more uncertainties, such as those due to climate change, energy markets, population pressures and ecosystem services. Diagnostic assessments of state-of-the-art multi-objective evolutionary algorithms (MOEAs) in support of EMODPS have highlighted that search time (or number of function evaluations) and auto-adaptive search are key features for success. Auto-adaptive MOEA search operators are themselves sensitive to having a sufficient number of function evaluations to learn successful strategies for exploring complex spaces and for escaping from local optima when stagnation is detected. Parallelizing MOEA search poses promising benefits for reducing the wall-clock time required to solve an application. Additionally, parallel search can facilitate more auto-adaptive algorithmic learning by enabling communication between parallel processes. This study explores how alternative parallelization strategies and stochastic sampling schemes serve to enhance the effectiveness and efficiency of the EMODPS framework. Our analysis focuses on the Lower Susquehanna River Basin (LSRB) system where multiple competing objectives for hydropower production, urban water supply, recreation and environmental flows need to be balanced. Our results provide guidance for balancing exploration, uncertainty, and computational demands when using the EMODPS framework to discover key tradeoffs within the LSRB system.