GCAM-Gym: a test bed for comparing algorithms that seek the best climate policy

Based on the GCAM integrated evaluation model, the environment was used to compare seven methods across 3,210 simulations; the CMA-ES evolutionary strategy emerged as the winner in all three scenarios.

Automatically seek a climate policy trajectory — how much to reduce emissions, when, at what cost — involves running heavy integrated assessment models like GCAM. Comparing research methods is difficult because they do not control the same levers, do not observe the same data and do not have the same simulation budget.

A common interface

Researchers, including Irina Rish, offer GCAM-Gym, an environment compatible with the Gymnasium library for strengthening learning. It makes it possible to evaluate a comprehensive policy as well as to steer the model step by step, and makes visible the different components of the objective. Seven search and control systems were compared in three configurations: achieving an objective at the lowest cost, arbitrating between cost and temperature, and feedback delayed from one step, with ten random runs each, or 3,210 GCAM evaluations attempted.

Among the four systems limited to 24 runs, the CMA-ES evolutionary strategy achieves the best results in all three configurations. The authors publish the full evaluation archive and analytical tools, allowing them to reproduce their comparisons without having to restart GCAM.