Simulating Equitable Waste System Transitions with ABM and Reinforcement Learning: Insights from Costa Rica

Costa Rica faces persistent disparities in recycling access and outcomes across its urban–rural divide. This study introduces the country’s first national-scale, multi-model framework integrating geospatial simulation, agent-based modeling (ABM), reinforcement learning (RL), and cost analysis to optimize waste management performance. The ABM employs machine learning–driven agents whose recycling decisions evolve based on spatial access, incentives, and demographic factors. A Random Forest classifier improves behavioral realism compared to static rules. Building on this, the RL module enables agents to iteratively learn optimal policies using spatial equity–shaped reward signals. Results show that while baseline ABM recycling rates vary between 28.9% and 36.7%, the RL-enhanced model raises national uptake to 84.3%, reduces backlog, and cuts cost-per-ton from USD 1,784 to USD 128. Spatial analysis reveals over 18,000 rural households remain over 50 km from a facility, underscoring infrastructure gaps. Monte Carlo simulations identify setup, processing, and transport as dominant cost drivers, while RL strategies show high resilience under sensitivity tests. This replicable framework supports data-driven, equity-centered planning for Sustainable Development Goals (SDGs), offering new insights for circular economy transitions in Costa Rica and similar Global South contexts.