Quality-of-Life-Driven Climate Adaptation Planning with Surrogate-Embedded Reinforcement Learning

Adapting cities to intensifying pluvial flooding is a sequential decision problem: adaptations must be allocated over multiple decades, under uncertain climate projections, and complementary with past interventions. Current planning paradigms typically optimise expected economic damages, which capture repair costs but not how flooding is experienced by individuals, i.e, a temporary, unevenly distributed loss of access to amenities and services. Evaluating such richer objectives requires chaining rainfall projections, flood modelling, transport network disruption, and accessibility-based welfare estimation, a pipeline far too slow to embed in any optimization or learning frameworks. To overcome this challenge, we propose a surrogate model-embedded reinforcement learning (RL) framework for long-term adaptation planning that optimises for individuals' quality of life. A learned surrogate model replaces the computationaly expensive accessibility and quality-of-life components, allowing an RL agent to learn what, when, and where to adapt under stochastic climate scenarios. Such surrogate-enabled IAM+RL frameworks let planners generate and stress-test long-term adaptation pathways against wellbeing outcomes directly. More broadly, learning the expensive links of an impact chain generalises beyond quality of life, enabling planners to optimise and assess more complex types of impacts rather than being computationally restricted.