Exploiting multi-objective reinforcement learning and explainable artificial intelligence to better navigate deep uncertainties in water supply infrastructure pathways

Dynamic and adaptive policy pathways frameworks are being increasingly applied to guide deeply uncertain water infrastructure investments and adaptation strategies in systems around the world. Evolutionary multi-objective reinforcement learning (eMORL) has direct value for advancing these frameworks by improving our ability to better represent complex state-actions dynamics across actors and timescales.  eMORL frameworks offer the potential to better understand the dynamics of state-aware actions that are contextually appropriate to the specific states of the world being experienced by system actors. However, the implications of the tradeoffs represented across alternative adaptive water supply investment policies pose nontrivial communication challenges. Investment pathways performance tradeoffs are typically communicated using highly aggregated metrics distilled to a single, expected value across actors and time. Here, this work addresses two main challenges. First, aggregated summary metrics do not capture the time-varying impacts of deeply uncertain (DU) factors on individual and system-wide performance and robustness. Second, aggregated summary metrics do not convey transparently state-action interdependencies between actors and performance objectives across time.

Our results address these challenges using a six-utility cooperative water supply infrastructure investment pathways example for the Research Triangle region in North Carolina. In our results, we contribute by-world, by-actor investment pathway diagnostics that clarify the consequential external deep uncertainties and state information feedbacks over time that strongly shape individuals’ adaptive actions. First, time-varying SHAP analysis clarifies the dynamics of which DU factors explain significant robustness shifts over time and across actors. Second, Information Theoretic Sensitivity Analysis identifies the key state variables that drive actions for each utility during specific periods of stress. In summary, our results can help decision-makers better understand how to navigate evolving vulnerabilities in their investment pathways and improve monitoring strategies to track changes consequential deep uncertainties over time.