Save Even More Energy with Cooperative AI

5G base stations
With cooperative AI (artificial intelligence), it is possible to simulate multi-agent exchanges in order to evaluate the suitability and feasibility of collaboration. At Orange, research is focused on applying this technology to negotiations between networks to optimize the energy efficiency and impact of their access networks.

These challenges are all the more critical given that RAN (Radio Access Network) infrastructure represents between 80% and 90% of the total energy consumed by a mobile carrier’s networks. Efforts are already being made to mitigate carriers’ energy footprints, for example, Orange is only using one frequency band at night when there is less traffic. However, other avenues can be explored to take this a step further.

A Cooperative Solution

For example, different carriers could work together to pool their resources, resulting in a significant reduction in the energy load of networks. In the scenario studied by Orange and presented at the Research and Innovation Exhibition, each carrier could alternate and host the others’ traffic for a one-night shift. That way, only one RAN would remain active, rather than two, three, four, or however many there are, depending on the number of carriers involved in the collaboration. The theory is promising on paper, but it must meet certain conditions in terms of fairness and incentive to contribute before it can be put into practice. Put simply: Will carriers want to take part? “To answer this question, we came up with some rules and launched cooperative trials,” explained Xavier Marjou, Network Research Engineer at Orange. “By experimenting with AI according to different sets of rules, it is possible to ensure that the proposed system of cooperation is reliable, fair and beneficial to all parties.”

Simulated Negotiations

Negotiations are simulated between carriers and we try to get everyone to get involved and take turns doing their “shift.” “We instantiated software agents representing the interests of each carrier, and the carriers learned to negotiate with each other over the course of the trials. It is a reinforcement-based learning model: The agent gets involved and performs an action toward an environment which, in return, feeds back the result of this action with an associated reward, positive or negative. By carrying out a series of actions, the agent ultimately finds the combination that is most favorable to them. In our use case, the agent/carrier offers to take a shift and in return receives a reward proportional to the amount of kWh saved. Over time, the agent realizes that certain negotiation sequences are more suitable than others and yield a profit, and they are therefore encouraged to continue taking part.”

AI learns the negotiating sequences that are favorable to them, and they are therefore encouraged to continue taking part and cooperating.

This application falls within the scope of cooperative AI, a relatively untapped sector due to significant computing power being required to be able to understand all combinatorial hypotheses in a multi-agent exchange.

The Best Rules to Encourage Cooperation

Many rounds of negotiations were simulated, producing various results. Certain rules are conducive to cooperation and a good level of alternation in shifts. Others, more biased, can lead an agent/carrier to exploit a loophole to the detriment of the other players. The system studied shows the diverse range of possible situations and proves that ethical cooperation is possible based on four evaluation metrics: efficiency (the amount of kWh saved), security (measures the risks for an agent when others do not cooperate), incentive to take part (encourages cooperation by showing the benefits) and fairness (each player saves the same amount of kWh). By modeling the interactions between carriers, it is possible to define the best negotiation parameters and limit exchanges to those strictly necessary, for optimal replication in “real life.”

The work carried out already demonstrates the benefits of having the largest number of carriers possible cooperating and pooling their resources, namely during periods of low activity.

Read also on Hello Future

Is AutoML the key to the democratization of machine learning?


Earthquake alerts: a convolutional neural network is monitoring the Earth’s gravity field


AI is facilitating the exploitation of the astronomical amount of space data


Behind the Scenes of AI — The Challenges and Methods of Explainability


IAlerting: Protecting Customers from Online Piracy


Using AI to Predict Diseases in Olive Groves

A woman in front of a computer analyzing electricity consumption data

Reducing mobile network power consumption with AI


When the evolution of species inspires algorithms