InvestESG: A Multi-agent Reinforcement Learning Benchmark for Studying Climate Investment as a Social Dilemma

We propose to create a multi-agent reinforcement learning (MARL) benchmark, InvestESG, to explore the impact of mandatory ESG (Environmental, Social, and Governance) disclosures on corporate climate investment behavior and market dynamics. By creating a MARL environment with companies and investors as agents, we will assess whether mandatory disclosure promotes corporate investment in reducing emissions, and how this is affected by investors’ preference for ESG efforts. We show via a Schelling diagram analysis that mitigating greenhouse gas emissions presents a social dilemma for various stakeholders such as firms, investors, and governments, where the benefits of reduced emissions are shared globally yet the cost are borne locally. We propose to first establish this the InvestESG social dilemma environment as a benchmark and challenge problem for MARL researchers, and subsequently to launch a competition for participants to submit solutions for designing regulations or incentives that achieve an overall reduction in global emissions. Building on the successful application of MARL in game-based social dilemmas and large-scale socio-economic issues, our benchmark provides a new window into the policy debate around mandatory ESG disclosure by enabling the analysis of the long-term interplay between firms and investors at a scale beyond existing studies in business and economics.