Abstract

In this thesis, a mean-field approach is used to find high-performance control strategies for multi-agent systems. The system consists of one leader and possibly many dynamically coupled followers, and all agents are affected by noise. The global objective of the multi-agent control system here is to achieve an agreement between the agents while minimizing coupled linear-quadratic cost functions for two cases: a disturbance-free system, and a system with disturbances. In the former case, the proposed solution under non-classical information structure is near-optimal, which converges to the optimal solution for a large number of followers. For the latter case, the problem is solved for three non-classical information structures, namely, mean-field sharing, partial mean-field sharing, and intermittent mean-field sharing. Using the minimax control technique, it is shown that the solution obtained for the first structure is a unique saddle-point strategy. On the other hand, it is proved that for the other two structures, the proposed solutions tend to the unique saddle-point strategy when the number of followers goes to infinity. The proposed strategies in both cases are linear, scalable and computationally efficient. The theoretical findings are verified by simulation results.