General rightsThis document is made available in accordance with publisher policies. Please cite only the published version using the reference above.

This study presents a novel approach which combines direct policy search and multi-objective evolutionary algorithms to solve high-dimensional state and control space water resources problems involving multiple, conflicting, and non-commensurable objectives. In such a multi-objective context, the use of universal function approximators is generally suggested to provide flexibility to the shape of the control policy. In this paper, we comparatively analyze Artificial Neural Networks (ANN) and Radial Basis Functions (RBF) under di↵erent sets of input to estimate their scalability to high-dimensional state space problems. The multi-purpose HoaBinh water reservoir in Vietnam, accounting for hydropower production and flood control, is used as a case study. Results show that the RBF policy parametrization is more e↵ective than the ANN one. In particular, the approximated Pareto front obtained with RBF control policies successfully explores the full tradeo↵ space between the two conflicting objectives, while the ANN solutions are often Pareto-dominated by the RBF ones.

No abstract

Multipurpose water systems are usually operated on a tradeoff of conflicting operating objectives. Under steady state climatic and socioeconomic conditions, such tradeoff is supposed to represent a fair and/or efficient preference. Extreme variability in external forcing might affect water operators' risk aversion and force a change in her/his preference. Properly accounting for these shifts is key to any rigorous retrospective assessment of the operator's behaviors, and to build descriptive models for projecting the future system evolution. In this study, we explore how the selection of different preferences is linked to variations in the external forcing. We argue that preference selection evolves according to recent, extreme variations in system performance: underperforming in one of the objectives pushes the preference toward the harmed objective. To test this assumption, we developed a rational procedure to simulate the operator's preference selection. We map this selection onto a multilateral negotiation, where multiple virtual agents independently optimize different objectives. The agents periodically negotiate a compromise policy for the operation of the system. Agents' attitudes in each negotiation step are determined by the recent system performance measured by the specific objective they maximize. We then propose a numerical model of preference dynamics that implements a concept from cognitive psychology, the availability bias. We test our modeling framework on a synthetic lake operated for flood control and water supply. Results show that our model successfully captures the operator's preference selection and dynamic evolution driven by extreme wet and dry situations.

No abstract