UMN Students Earn First Place in Global AI Telco Troubleshooting Competition

A Department of Computer Science & Engineering team claimed first place in track three of the Artificial Intelligence Telco Troubleshooting Challenge, a global competition aimed at accelerating innovation in using large language models (LLMs) for root cause analysis in telecommunication networks. Their winning solution earned the group an $8,500 prize and an invitation to present their work at the Mobile World Congress (MWC) in Barcelona – the largest annual mobile industry event. 

The winning team was made up of Wei Ye, a recent University of Minnesota computer science Ph.D. graduate and current artificial intelligence (AI) researcher at Accenture’s Center for Advanced AI, Vivian Tsang, a recent computer science master’s graduate and a current software engineer at Certusoft, and Gazi Fahim Abrar, a computer science BS student. The three worked together in the Computer Networking, Mobile, and AI Research Lab, which focuses on advancing the reliability, availability, and security of modern communication networks. 

When presented with the opportunity by their lab advisor, Professor Zhi-Li Zhang, the group felt they were a good fit for the competition. 

“We felt it would be a good match with what we are each good at,” Ye said. “Our team has extensive expertise in applying AI to network-related problems, developing tools, and publishing research, so we saw this as a great opportunity.”

The competition's challenge was to improve the foundation models capable of diagnosing failures in telecommunication networks, a task with real‑world impact, the team explained. 

If a base station is broken, the language model can identify or analyze what the root reason is and provide instructions to a human operator,” Ye said. “A human operator can ask the AI agent why the performance is going down, and the agent can tell them why.”

Teams were given a public dataset of cell‑tower and user information to build and test their models during the competition. Final results were determined using a separate private dataset. 

Over two months, the team developed their solution, though it did not score highly when tested on the public dataset. 

“We thought, okay, we participated,” Abrar said after submitting their solution. “Somehow, surprisingly, our solution performed great on the private data set, and we won the competition.”

Their solution showed that strong performance on a familiar dataset doesn’t necessarily reflect how well a model will perform on new, unseen data.

“In the AI domain, when we train a model using a specific data set, it’s very easy to overfit and over-memorize the training data set, causing the poor generalization ability to the test data set,” Ye said. 

For the team, winning affirmed that their work was headed in the right direction.

“It’s motivating,” Ye said. “It recognizes that our research is not alone. There are thousands of other people participating in the competition. It shows our topic has value and that our research is not a random topic.”

In March, Tsang represented the team at MWC, where she met the tournament organizers and connected with other winning teams.

“It was pretty exciting,” Tsang said. “It was fun to see what kind of work is being done at the forefront of the field and what other people are working on.”

Looking ahead, the team plans to improve their solution through deeper fine tuning and by incorporating temporal patterns in the data. They say their approach is especially promising because of its small size, making it deployable on a wide range of devices with modest GPU resources.

The team also expressed gratitude for Professor Zhang’s support. 

“We wanted to give him special thanks for all his support,” Ye said. “He provided the facilities, the connections, and the opportunity to take part in this competition.”