Beyond Expected Value: PPO with Distributional Critics for Systematic Conservation Planning

Systematic conservation planning must allocate limited protection budgets across space and time under deep ecological and climatic uncertainty. Yet, reinforcement learning (RL) approaches optimise only expected outcomes and remain blind to worst-case scenarios. We extend CAPTAIN, an RL framework for conservation planning, replacing its gradient-free Evolution Strategies (ES) optimiser with Proximal Policy Optimisation (PPO) via Plackett-Luce action sampling, and replacing PPO's scalar critic with an Implicit Quantile Network (IQN) that models the full distribution of ecological outcomes. On a New Zealand marine biodiversity dataset, PPO's gradient-based formulation reaches conservation outcomes comparable to or better than ES at a fraction of the training cost, and uniquely enables a distributional (IQN) critic that improves outcomes further still. A matched-granularity ablation shows decision granularity, not the optimiser, drives most of the raw ecological gain, so we frame the pipeline's real contribution as an extensible foundation for risk-sensitive conservation planning rather than the specific outcome numbers reported here. The method consumes standard species distribution model inputs and is not tied to any one region by construction.