Researchers at the University of Illinois Grainger College of Engineering have created a technique that can use artificial intelligence to make many agents cooperate (AI). 

The approach uses multi-agent reinforcement learning, a subset of AI. Individual agents, including robots and drones, can frequently cooperate and do a task when communication channels are open. However, if the appropriate hardware is not present or signals are blocked, devices may hinder communication between them. By identifying each agent when it contributes to the team's goal, the new method fills this gap using AI.

Huy Tran, an aeronautical engineer at Illinois, noted that it is more straightforward when agents can communicate. "But we wanted to do this, so they don't talk to each other. We also paid attention to situations where it wasn't clear what the agents' different roles or jobs should be."

According to Tran, the situation is complicated since it is unclear what one agent should perform relative to another. He stated that the issue was to finish a task collectively over time. Tran and his coworkers employed machine learning and developed a utility function that notifies the agent when contributing to the team's efforts.

Tran noted that it is difficult to discern who contributed to a team goal's success. This void is by the machine learning technology, which identifies the individual agent when it contributes to the team objective. "From a sporting perspective, one soccer player may score, but we also want to know about activities by other teammates, such as assists, that lead to the goal. It is difficult to comprehend these delayed impacts," he continued.

Image source: The Grainger College of Engineering

The algorithms the researchers used can also spot when an agent or robot is acting in a way that isn't helpful to the result. The robot chose to do something that wasn't helpful to the result, not necessarily anything that was bad. They used simulated games like Capture the Flag and StarCraft, a well-known computer game, to evaluate their algorithms. Furthermore, StarCraft can be a little more unexpected, so we were happy to learn that our approach also worked well there.

Conclusion

According to Tran, this algorithm is relevant to many real-world scenarios, including military surveillance, robot collaboration in a warehouse, traffic signal management, delivery coordination by autonomous vehicles, and grid control.

Furthermore, Tran stated that Seung Hyun Kim developed most of the idea's theory while an undergraduate studying mechanical engineering, while Neale Van Stralen assisted with its implementation. In addition, Tran and Girish Chowdhary served as advisors for both students. Recent presentation to the AI community at the peer-reviewed Autonomous Agents and Multi-Agent Systems conference.

Image source: Unsplash