Accounting for human implementation uncertainties in the discovery of robust water supply infrastructure pathways using evolutionary multi-objective reinforcement learning.

Urban water utilities face the combined pressures stemming from evolving drought extremes, increasing demands, and financial constraints, prompting a growing interest in regional cooperative dynamic and adaptive infrastructure investment pathway strategies. Theoretically, these strategies promise improved resource efficiency by realizing economies of scale, adding flexibility for achieving improved supply reliability, and, ideally, limiting individual and collective financial risks. However, prior work has shown that implementation uncertainty in regional partners’ cooperative actions, characterized by modest deviations from a prescribed set of Pareto-approximate actions, can drive counterparty risks and potentially exacerbate collaborating actors’ vulnerabilities to deeply uncertain supply and financial challenges. To address this challenge, we contribute the Deeply Uncertain Pathways for Implementation Uncertainty (DU Pathways IU) framework, an evolutionary multi-objective reinforcement learning (eMORL) approach that accounts for human-driven implementation uncertainty when optimizing for regional cooperative water supply management and planning pathway strategies that remain robust to external socio-economic uncertainties and drought extremes.

In this work, we demonstrate that the DU Pathways IU approach yields a broader range of regional water supply pathway strategies that more fully utilize the full suite of cooperative management and planning actions available to regional actors. This broader set of highly cooperative pathway strategies exhibit more controlled supply and financial performance degradation when stress-tested under implementation uncertainties (i.e., perturbations to pathway strategies’ decision variables). In addition to remaining stable in the face of unexpected deviations from the recommended set of regional cooperative actions, these strategies achieve higher robustness across all regional actors. Further sensitivity analysis reveals that highly cooperative pathway strategies experience reduced sensitivity to perturbations to other actors’ actions. Consequently, cooperating utilities have more control over their individual performance and reduced uncertainty when assessing the needed timing and prioritization of future infrastructure investments. Overall, this work facilitates the discovery of highly cooperative regional water supply planning and management pathway strategies that remain stable under implementation uncertainty. It is broadly applicable to water utility managers who seek improved transparency into how modest perturbations in cooperative actions drive potential performance conflicts across multiple actors implementing both individual and cooperative actions in a regional system.