Supervisory decision and control of large-scale multi-agent systems.

摘要

This dissertation addresses the general problem of optimally controlling a large number of reactive autonomous agents under certain performance constraints. The proposed solution is to divide the control architecture into layers. In addition to incorporating the local interactions amongst the agents and with the environment, the lower layer encompasses signal processing tools from probabilistic finite state automata (PFSA) based models. On the other hand, the upper layer performs high level tasks such as planning, optimization and learning. The upper layer incorporates probabilistic supervisory decision and control on a lower dimensional manifold of the configuration space of the agents. The entire team of agents is modeled as a PFSA and control over the team of agents is exerted by varying the probabilities of state transitions in a continuous domain. For homogenous agents, complexity of the proposed algorithm is independent of the number of agents; hence, this decision and control algorithm is applicable to swarms of arbitrary size. Furthermore, the supervisory controller makes use of the asymmetric broadcast control paradigm, where all the agents receive identical instructions, although individual agents may act differently depending on their current states. The proposed algorithm has been validated on a simulation test bed of an underwater repositionable sensor network.