With Evolutionary AI, Scientists Find Hidden Keys for Better Land Use

Using global land use and carbon storage data from the past 175 years, researchers at The University of Texas at Austin and Cognizant AI Lab have trained an artificial intelligence system to develop optimal environmental policy solutions that can advance global sustainability initiatives of the United Nations. The AI tool effectively balances various complex trade-offs to recommend ways of maximizing carbon storage, minimizing economic disruptions and helping improve the environment and people’s everyday lives, according to a paper published today in the journal Environmental Data Science. 

The project is among the first applications of the UN-backed Project Resilience, a team of scientists and experts working to tackle global decision-augmentation problems—including ambitious sustainable development goals this decade—through part of a broader effort called AI for Good. University of Texas at Austin computer scientist Risto Miikkulainen, who helped launch Project Resilience, believes the new AI approach, initially focused on land use, can address an even larger set of challenges, from infectious diseases to food insecurity, with artificial intelligence potentially discovering better solutions than humans.

“There’s always an outcome you want to optimize for, but there’s always a cost,” he said. Amid all of the trade-offs, AI can home in on unexpected pathways to desirable outcomes at various costs, helping leaders selectively pick battles and yield better results.

The secret sauce of the researchers’ system is evolutionary AI. Inspired by the process of natural selection in biological systems, this computational approach starts with a few dozen policy scenarios and predicts how each scenario will impact various economic and environmental costs. Then, like a digital version of survival of the fittest, policy combinations that don’t balance the trade-offs well are killed off, while the best ones are allowed to reproduce, giving rise to hybrid offspring. Random mutations also are sprinkled in to help the system explore novel combinations faster. The process then repeats, winnowing poor performers and keeping the best, across hundreds or thousands of scenarios. Like biological evolution, the “generations” of scenarios become ever-more optimized for a set of priorities.

The team used two tools—a recently released set of global land use data going back centuries and a model that correlates land use with carbon fluxes. First, they used this data to train a prediction model to correlate location, land use and carbon over time. Second, they developed a prescription model to help decision makers find optimal land-use strategies to reduce climate change. 

The AI system’s recommendations sometimes surprised the team. Although forests are known to be good at storing carbon, the AI prescription model offered a more nuanced approach than converting as much land as possible into forests, regardless of location. For example, it found that replacing crop land with forest is much more effective than replacing range land (which includes deserts and grasslands). Also, generally, the same land use change at one latitude didn’t yield the same benefits as at another latitude. Ultimately, the system recommended that larger changes should be allocated to locations where it mattered more; in essence, it’s more effective to pick your battles.

“You can obviously destroy everything and plant forests, and that would help mitigate climate change,” said Daniel Young, a researcher at Cognizant AI Lab and a Ph.D. student at UT Austin. “But we would have destroyed rare habitats and our food supply and cities. So we need to find a balance and be smart about where we make changes.”

The researchers have turned their model into an interactive tool that decision makers like legislators can use to explore how incentives, such as tax credits for landowners, would be likely to alter land use and reduce carbon.

Land use activities, including agriculture and forestry are estimated to be responsible for nearly a quarter of all human-caused greenhouse gas emissions. Experts believe smart land use changes will be needed to reduce the amount of carbon in the air and thereby slow climate change. According to Miikkulainen and Young, AI offers options that people, businesses and governments otherwise resistant to change may find easier to accept.

An earlier version of the paper was presented at a major machine learning and computational neuroscience conference, NeurIPS, where it won the “Best Pathway to Impact” award at the Climate Change workshop.

The other authors on the paper are Olivier Francon, Elliot Meyerson, Clemens Schwingshackl, Jakob Bieker, Hugo Cunha and Babak Hodjat.

Expanding Reach

The research team is working with the nonprofit Climate Interactive to layer their evolutionary AI approach on top of an existing online climate simulator, called En-Roads. En-Roads currently enables decision makers, teachers, businesses, journalists and the general public to plug in a host of policy choices — such as electrifying transport or pricing carbon — and then see what impact that has on future global temperatures. Climate Interactive regularly facilitates interactive workshops using their simulator with business leaders and government decision makers, hosting interactive sessions, as well, in classrooms, community groups and beyond. 

En-Roads involves users adjusting different policy levers to predict future outcomes, but in place of such a trial-and-error-based approach, Miikkulainen’s team is working towards allowing users to select a target for future global temperature — say, no more than 1.5 degrees warmer than pre-industrial times — with the AI model taking charge of optimizing a set of policies to get there, based on the users’ preferred trade-offs. They’ve created an add-on tool called Decision Making for Climate change that applies their evolutionary AI approach to En-Roads.

“They are just amazed at what you can do,” Miikkulainen said. “They didn’t even think that it could be possible to automatically make recommendations that meet certain goals.”

Previously, the same team involved with the sustainability efforts also applied evolutionary AI in exploring options to combat the spread of COVID-19.

“We can’t actually try out all the different possible combinations of non-pharmaceutical interventions in a pandemic,” Miikkulainen said. “But we have an AI system that can try out many different combinations and discover good policies. The system discovered early on surprising things we did not anticipate,” such as the importance of timing lockdowns as early as possible.

In summer 2021, a year and a half into the pandemic, the Icelandic government asked Miikkulainen and his team to advise on school openings for the coming school year. The researchers customized their evolutionary models with additional data specific to that country and made recommendations that helped decision makers there set effective policies. They also helped organize an XPRIZE competition for other teams around the world developing their own AI systems capable of prescribing intervention and mitigation measures for COVID-19 and future pandemics.

All these domains—land-use optimization, climate change mitigation, and pandemic intervention—share the same challenge of modeling the world and optimizing the response. There are many other problems that could be addressed in the same manner. Thus, the team has open-sourced the fundamental technology in the Project Resilience platform. The idea is to allow anyone to get involved: identify opportunities, build their own models, and propose solutions, for instance to address UN’s sustainable development goals, and others. 

Seeing all of the ways that teams, many with support and encouragement from the UN, are effectively using AI to aid in complex decision making has made Miikkulainen more hopeful for the future.

“I’m seeing a lot of growth and excitement around the AI for Good movement,” says Miikkulainen. “Researchers, journalists and governments are finding new opportunities to use AI to make the world a better place. It’s just amazing how many people see that and are actually working on it.”