_{Conceptual struture of a multiagent system}

Highlights:

❀ Assumed that the control constraint set is finite in order to argue about the computational efficiency of the agent-by-agent rollout algorithm.

❀ Proposed new autonomous multiagent rollout schemes for both finite and infinite horizon problems. The idea is to use a precomputed signaling policy, which embodies sufficient agent coordination to obviate the need for interagent communication during the on-line implementation of the algorithm.

❀ Finally mention that the idea of agent-by-agent rollout also applies within the context of challenging deterministic discrete/combinatorial optimization problems, which involve constraints that couple the controls of different stages.

推荐阅读：

美国工程院院士Dimitri P. Bertsekas: 强化学习及最优控制(71页PPT)

美国工程院院士Dimitri P. Bertsekas: 基于特征的聚合与深度强化学习