What is it about?
Many homes with electric cars could earn money by selling stored battery power back to the electricity grid when prices are high, then recharging when prices are low. But deciding exactly when to do this is hard. This project used an AI language model, similar to ChatGPT, to write the computer code that makes these decisions. The AI tested its own code, learned from the results, and improved itself automatically. The final code was just 15 lines long, easy for anyone to read, yet it earned 18% more money than rules written by human experts, and even discovered smart strategies on its own.
Featured Image
Photo by dcbel on Unsplash
Why is it important?
As more households adopt electric vehicles and home solar panels, the way we control these batteries will matter more and more, both for household savings and for the stability of the wider electricity grid. Right now, the tools available force a difficult choice. Advanced systems can perform well but work like a black box, so nobody can check why they make the decisions they do. Simple, human-written rules can be understood by anyone but often fail when electricity markets behave unexpectedly. This work shows a third path is possible. By using an AI language model to write and refine its own control code, we get a system that performs well, adapts to changing conditions, and remains simple enough for a person to read, question, and trust. This is especially important for household energy technology, where families, electricians, and regulators all need to be able to understand and approve how a system behaves before it is allowed to control something as important as home power and transport. The approach is also cheap and practical. The entire improvement process cost about 20 cents in computing, and the resulting code can run on ordinary hardware, without needing powerful computer chips. This makes the method realistic for real homes, not just research laboratories, and offers a template for building trustworthy AI systems in other areas of everyday life where transparency matters just as much as performance.
Perspectives
Working on this project changed how I think about using AI in research. Early on I was skeptical that a language model could produce anything beyond a rough first draft of a control policy. What I found instead is that with the right feedback loop, it can be pushed toward genuinely strong, well reasoned solutions. It shifted my view of these models from a shortcut to a real collaborator in the design process.
Vishesh Purnananda
University of Adelaide
Read the Original
This page is a summary of: Evolving LLM-Derived Control Policies for Residential EV Charging and Vehicle-to-Grid Energy Optimization, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3795101.3805325.
You can read the full text:
Resources
Contributors
The following have contributed to this page
