What is it about?
This paper shows how a neural network can replace a slower climate model inside an AI system for testing climate policies. The AI system has multiple agents, such as countries or regions, that learn which climate actions to take over time. The neural network is trained to imitate CICERO-SCM, a simple climate model that predicts global temperature change from emissions of several gases. It gives almost the same temperature results, but runs much faster.
Featured Image
Photo by NOAA on Unsplash
Why is it important?
Climate-policy AI experiments need to run climate models millions of times. Even simple climate models can be too slow for this. By using a fast and accurate surrogate model, researchers can test richer climate-policy scenarios with more agents, more greenhouse gases, and more realistic climate responses. In our experiments, the surrogate was about 1000 times faster for climate predictions and made reinforcement-learning training more than 100 times faster, while still leading to the same learned policies as the original climate model.
Perspectives
I developed this work to make climate-policy reinforcement learning more realistic without making it computationally impossible. The main motivation was to remove the climate-model bottleneck, so researchers can study richer policy settings with multiple agents and multiple greenhouse gases. I see this as a step toward more scalable experiments on climate cooperation, where AI agents can explore policy choices while still using scientifically meaningful climate responses.
Oskar Bohn Lassen
Danmarks Tekniske Universitet
Read the Original
This page is a summary of: Climate Surrogates for Scalable Multi-Agent Reinforcement Learning: A Case Study with CICERO-SCM, International Foundation for Autonomous Agents and Multiagent Systems,
DOI: 10.65109/rjbj9974.
You can read the full text:
Contributors
The following have contributed to this page
