Projected gradient-tracking in multi-cluster games and its application to power management

We are concerned with a distributed approach to solve multi-cluster games arising in multi-agent systems. In such games, agents are separated into distinct clusters. The agents belonging to the same cluster cooperate with each other to achieve a common cluster goal while a non-cooperative game is played between the clusters. To be able to deal with the sparsity of information, as each agent only knows a specific part of the problem, we combine gradient-tracking and consensus methods for information distribution into an algorithm that can solve both the cooperative and non-cooperative problem in a single run. The constraints of the problem are taken into account by the corresponding projection operators and linear convergence is proven given an appropriate constant step size. The algorithm is applied to a day-ahead power management problem, posed as a multi-cluster game, and its efficiency is demonstrated by simulations.