Agents of Change

“A wildfire could pop up in a place you’re not expecting—or the winds could shift, and now it’s heading toward a population center,” Eck says. “The world around us is so dynamic: How do we model that, account for it, and make decisions in the presence of that?”

In artificial intelligence terms, these robots are known as autonomous agents and possess humanlike qualities. “An autonomous agent is an AI that acts independently on its own,” Eck says. “It gathers information from the world, makes its own decisions about how to accomplish its goals and tasks, and then takes actions to physically change the world.”

When the robots are figuring out how to put out wildfires, cooperation is key, which is where Eck’s research comes in: He studies the social side of artificial intelligence, where multiple autonomous agents gather in what’s called a multiagent system. 

“We want the robots to come up with strategies to fight fires together, to put everything out as fast as possible,” he says. “A big part of this social side is modeling what the other AIs are doing, predicting their actions, and then trying to cooperate with them. If they all choose to fight individual fires, then they’re not going to be nearly as strong as if they do things together.”

For the robots, this is easier said than done. These multiagent systems are in turn operating within a complex open agent system that’s always changing. It’s likely the robots aren’t setting up rules or coordinating actions ahead of time; as a result, they must predict all potential scenarios they might face in a wildfire. Autonomous agents also shift in and out of these open systems, requiring the robots to first predict who’s around and then try to figure out how to work with that group. 

Complicating matters further is that these open agent systems have task openness, which means the set of tasks someone is trying to accomplish changes over time. “The wildfire-fighting robots might get picked up and moved to a whole new area,” Eck explains. “In that case, they’d have to reorient themselves and say, ‘OK, I’ve got different fires I need to fight now. How does that change my behaviors?’”

Eck notes these open agent systems also possess type openness, where the capabilities of the agents change over time. “Maybe the robots get damaged and can’t fight fires as well anymore. How does that change their decision making? How do you continue working with someone who has new capabilities?”

To date, Eck and his students are doing this research via computer simulations of wildfires, based on software that approximates how the real world works using either numbers or a visualization. While artificial intelligence is at the core of this research, Eck’s lab uses human insights to inform hypotheses. For example, they incorporated information about how fires spread and under what conditions from past simulations built by wildfire domain experts.

“We use human thinking as inspiration and try to imagine, ‘If I were to tackle this problem, what would I do?’” he says. “Afterward, it’s fun to see how the AI ends up solving it. It’s inspired by humans to begin with, but it might come up with entirely new ideas and do it in a different way that people hadn’t thought of before.” 

Eck isn’t building robots and sending them into the field to fight wildfires yet. The decision-making abilities of these autonomous agents aren’t quite fast enough for the unpredictable nature of the real world. Eck stresses that these multiagent systems are enormously complex and challenging to scale up. 

“It’s one thing for a robot to decide independent of everybody else, ‘What fire do I want to fight?’” he says. “But once you have to start modeling everybody else, the more neighbors that we have, the more time I have to allocate to predicting for each one of them. That slows down my own decision-making.”

Eck has received two National Science Foundation (NSF) research grants to study open agent systems, leading to multiple publications. Along with collaborators at the University of Georgia and University of Nebraska, he published a paper with the 2020 AAAI Conference on Artificial Intelligence. Eck has also published work focused on decision-making in open systems, including a paper at the 2022 UAI Conference on Uncertainty in Artificial Intelligence and a 2023 article in AI Magazine.

He’s also looking at other applications of these open agent concepts, including decision-making around dynamic ride sharing of autonomous cars; supporting business managers over time as they gain new responsibilities within complex organizations; and coordinating behaviors of cybersecurity agents protecting critical infrastructure.

“Much of AI is building on previous solutions,” Eck says. “You start by tackling simpler problems, and then you can get harder and harder and harder as you go along. We’re trying to model even more complicated environments to get closer to what the real-world situation is to make these problems easier to solve.”