Climate change mitigation requires balancing economic growth against environmental targets across diverse regions with competing priorities. Integrated Assessment Models (IAMs) are widely used to inform climate policy, yet existing Reinforcement Learning (RL) extensions rely on the homogeneity assumption — treating world regions as identical decision-makers. This misrepresents the diversity of economic conditions and climate vulnerabilities that characterise real-world global systems. This paper extends the JUSTICE Multi-Objective Multi-Agent RL (MOMARL) framework by introducing heterogeneous agents, enabling each region to maintain an independent policy. We propose two algorithms — MOHAPPO and MOHASAC — and evaluate them against the homogeneous baseline MOMAPPO. Heterogeneous algorithms achieve higher Hypervolume and Expected Utility, produce more diverse Pareto fronts, and identify more aggressive emission mitigation strategies. Regional analysis reveals that the distribution of mitigation burdens varies across algorithms and policy types, with some configurations placing disproportionate mitigation effort on vulnerable regions. We demonstrate how explicitly encoding equity in the reward structure — via utilitarian versus prioritarian welfare objectives — can shift the distribution of burdens across regions, reinforcing the need for mindful optimisation design. These findings underscore that
algorithmic design choices in RL-based climate modelling are not value-neutral, and argue for heterogeneous, equity-aware frameworks as essential in this domain. ... Read more