13:31, 14 March 2026

AI Assigned to Manage Network Traffic in Russia

Researchers have introduced a multi-agent algorithm that reduces data exchange between agents and reallocates traffic flows faster when network load changes.

Researchers at Moscow State University have developed a new algorithm for managing traffic in telecommunications networks and data centers. The study was published in the scientific journal Mathematics.

Traffic engineering helps distribute data flows between network nodes and prevent congestion. The rapid growth of internet traffic and the rising number of connected devices are making this task increasingly complex. Classical optimization algorithms often respond poorly to the dynamic and unpredictable workloads typical of modern networks.

As a result, researchers have been exploring multi-agent systems. In these systems, independent software agents monitor the network, make local decisions, and coordinate their actions with one another.

MAROH Algorithm

In earlier work, the team at the MSU AI Center introduced an algorithm called MAROH. It combines multi-agent optimization with machine learning.

The method showed higher efficiency than widely used load-balancing algorithms, including ECMP and UCMP. However, the system required intensive information exchange between agents. Every time traffic changed, the agents recalculated flow distribution parameters from scratch.

The new version of the algorithm changes the decision-making architecture. Developers added a dual-loop mechanism designed to reduce load on the network’s control layer.

Algorithm With “Fast” and “Slow” Thinking

The developers drew inspiration from research by Nobel Prize–winning economist Daniel Kahneman. He described two modes of human thinking: one that works quickly using accumulated experience, and another that activates in complex or unfamiliar situations.

Agents in the new system follow a similar logic. If the network state is familiar, they immediately apply a previously learned strategy. When a new traffic configuration appears, the agent launches a deeper analysis and coordinates with other agents in the system.

“We started from the idea that in distributed systems the coordination process itself can become a source of additional load. The dual-loop decision-making model allows an agent to make faster decisions in familiar states and rely on more complex analysis—requiring interaction with other agents—only when necessary,” said Evgeny Stepanov.

Reducing Network Exchanges

Experiments showed a significant effect from the new architecture. Depending on the algorithm’s parameters, the number of exchanges between agents decreases by 80–96 percent compared with conventional methods.

At the same time, the system maintains efficient traffic balancing and quickly reaches a stable distribution of data flows. This is particularly important for infrastructures with heavy workloads and constantly changing network activity.

“The idea behind the proposed method was inspired by human cognition. But implementing it as an intelligent agent capable of accumulating experience and forming a kind of intuition required deep mathematical foundations and the use of neural network technologies from several different classes. As a result, we were able to reduce unproductive network load and speed up decision-making without complicating the network architecture,” said Ruslan Smelyansky.

Science and new technologies