Poster: Green Network Optimization with Multi-Agent Reinforcement Learning: Work in Progress

This paper proposes a distributed Green Network Optimization (GNO) with Multi-Agent Reinforcement Learning (MARL). GNO-MARL combines: (a) Network Utility Maximization (NUM) based network management, given the capacity constraints, on a “fast” time scale, and (b) network dimensioning which balances the maximal network utility with the “carbon cost” of the transmissions on “slower” time scales. Inherent non-convexity of the GNO due to dominance of the idle power in network devices, including routers, makes exact solution of the GNO even by centralized algorithms for realistic size networks intractable. It is known that distributed algorithms may introduce additional to the duality gap loss in the performance due to oscillatory instability. We advocate that this additional loss may be significantly reduced or even eliminated with MARL. This possibility is based on properties of Lagrange multipliers and suggested by known results on reducing the corresponding additional loss in network performance.