The Zambezi River Basin (ZRB) is a critical resource for Southern Africa, supporting hydropower production, livelihoods, food security and ecosystems. With increasing freshwater scarcity and climate change induced droughts and floods in the ZRB, water allocation is increasingly critical, especially to those already carrying the burdens of climate change without reaping the profits of economic development. Effective management of this transboundary water system requires balancing competing objectives such as economic efficiency, social equity, and environmental sustainability.
In light of the DAFNE project, funded by the EU to create a Decision-Analytic Framework (DAF), an Evolutionary Multi Objective Direct Policy Search (EMODPS) framework was applied to the ZRB. EMODPS models combine Direct Policy Search (DPS) with Multi-Objective Evolutionary Algorithms (MOEA) to process complex simulations and continuously optimize for sequential decisions. The ZRB EMODPS model was created to identify the Pareto-optimal release policies for the five hydropower reservoirs and eight irrigation districts in the river basin. In the modelling process, there was a lack of consideration for distributive justice. In the baseline configuration, the five reservoirs were aggregated into one hydropower objective and the eight irrigation districts in the system were aggregated into one irrigation objective. The environmental flow at the Zambezi Delta constituted the third objective for the initial optimization.
This research disaggregates the hydropower and irrigation objectives to analyse what the effects are on the optimal release policies, particularly for smaller irrigation districts and reservoirs. The research question is: How does the disaggregation of objectives influence the Pareto space for an EMODPS simulation-optimization model? Four levels of aggregation were optimized: the baseline configuration with three objectives, the irrigation case with 11 objectives (including an individual objective for each irrigation district), the hydropower case with eight objectives (including the five hydropower reservoirs as objectives) and the full case with 16 objectives in total where the three baseline objectives are complemented with one objective for each irrigation district and hydropower reservoir. The Pareto set of the four different problem framings is visualized and analysed to conduct a comparison between the levels of aggregation.
Higher levels of aggregation may limit the insights provided by the Pareto front and increase the risk of further burdening marginalized groups. The initial hypothesis was that smaller irrigation districts and hydropower reservoirs would benefit from being considered as individual objectives. However, this hypothesis was not confirmed. The baseline aggregation of three objectives yielded better results for the total hydropower and irrigation deficits, even for the smaller districts and reservoirs.
The results reveal that disaggregation provides a more nuanced understanding of trade-offs but increases computational demands and complexity. The increased number of variables and constraints decreased the efficiency of the Generational Borg algorithm, making the study less feasible. Many-objective optimizations with more than 10 objectives pushed computational limits, displayed unexpected convergence behaviour, and posed challenges in presenting and interpreting large amounts of data. More sophisticated algorithms may better handle the consequences and limitations of objective aggregation in EMODPS models. This research highlights the trade-offs between equity and efficiency in water resource management and provides insights into the possibilities of disaggregating objectives for more just and precise policy-making.

Forest BORG, a novel framework using the BORG multi-objective evolutionary algorithm, optimizes policy trees for socio-economic systems. Tested on real-world case studies, Folsom lake and RICE IAM, it addresses multi-objective challenges under deep uncertainty, offering effective and explainable decision support for climate change policies.

The 'Forest BORG' framework, based on the BORG multi-objective evolutionary algorithm (MOEA), is introduced as a method for optimizing policy trees in socio-economic systems. This framework accommodates tree-structured decision variables through specialized evolutionary operators, enabling multi-objective and explainable decision support. Tested on the Folsom lake model and the Regional Integrated model of Climate and the Economy (RICE) Integrated Assessment Model (IAM), Forest BORG addresses the challenges of optimizing multi-objective problems under deep uncertainty, primarily driven by climate change.

The Folsom lake model and RICE IAM serve as suitable case studies due to their reflection of real-life problems affected by climate change and involving multiple stakeholders. The XLRM framework is employed to operationalize the conceptual models of these case studies, categorizing uncertainties (X), levers (L), relations (R), and metrics (M).

Positioned between current MOEA capabilities and the need for explainable policies in real-world multi-objective problems under deep uncertainty, Forest BORG undergoes rigorous testing. Analyses include run-time dynamics, Forest BORG specific evolutionary operator dynamics, controllability, exposed trade-offs, and highlighted policy trees.

The framework employs seven mutation operators, crossover with prune and bloat control operations as the recombination operator, and an adaptive tournament size function as the selection operator. The performance and behavior of Forest BORG are evaluated through the five analyses mentioned.

Run-time dynamics demonstrate the algorithm's effective convergence based on hypervolume. Search operator dynamics investigate novel mutation operators, showcasing their contribution. Controllability analysis explores the impact of hyperparameter settings on algorithm performance. Exposed trade-offs are visualized through pareto-optimal policy alternatives, emphasizing the selection of compromise policies. The highlighted policy trees provide insight into actionable decisions based on different objectives.

Forest BORG proves effective, efficient, and reliable across case studies. Comparative analysis with the original policy optimization tree (POT) algorithm reinforces its effectiveness. Although the expected shift from subtree to point mutations was not observed in all cases, the search operators enhance overall search capability. Hyperparameters are case study-dependent, with modest maximum depths favored to avoid overfitting.

Forest BORG successfully produces pareto fronts for each case study, revealing inherent trade-offs. The selected policy trees align with performance metrics, providing valuable insights for decision-makers. However, some results exhibit counter-intuitive behavior, notably in the RICE model trade-offs and small peaks in run-time analysis, attributed to epsilon values.

The research suggests further exploration of other IAMs for a more realistic representation of reality. Forest BORG's societal contribution lies in its development and application to real-world case studies, expanding the range of multi-objective decision-making tools for socio-environmental systems. The implications extend to addressing climate change through policy development, emphasizing the potential for broader applications beyond the tested case studies.