Distributed Control of Cooperative Multi-Agent Systems: Combined Top-down and Bottom-up Design
Abstract
Cooperative multi-agent systems are useful for many applications including coordinated surveillance, automated warehouse, traffic management, and manufacturing systems. A key issue in these systems is to design distributed coordination and control mechanisms so that certain desirable global specifications are met. Existing methods in multi-agent system control can be roughly divided into two categories: bottom-up and top-down design. Bottom-up methods, such as the consensus-based approaches, build upon pre-designed local coordination rules and let collective global behavior emerge. Hence, bottom-up approaches can scale well, but usually lack performance guarantees, as emergent collective behavior is hard to predict, none to speak of control. On the other hand, top-down design bases on a "divide-and-conquer" idea and reduce the multi-agent design to individual agents local synthesis with respect to decomposed local utilities or specifications. Although top-down design can provide performance guarantee, it relies on an abstracted quotient model of the agents in a finitely partitioned environment, which makes it hard to deal with complex physical dynamics or uncertain environments. We propose to combine top-down and bottom-up design methods so to leverage advantages from both. First, we investigate the mission decomposition problem that is the key issue in the top-down design of multi-agent systems. Second, from the bottom-up design perspective. we propose to adopt a passivity based design method to explore the composition of basic dynamic motion primitives. Finally, we will investigate how to combine our results into a unified framework. For this, we propose two complementary approaches. In the first approach, to achieve task specifications, we propose to encode the specifications into integer or mixed-integer constraints and use constraint solvers to find feasible solutions. The second approach builds on composability of passivity as the passivity is preserved when systems are interconnected in parallel or in feedback configurations. Thus, if we could certify that the behavior of the individual agents would satisfy some passivity indices or dissipativity inequalities, we could safely compose them. The significance of the proposed research centers on the fundamental question essential for building distributed cooperative multi-agent systems that can function robustly and reliably in unknown and dynamic environments. The proposed research, if successful, removes a major barrier in current practices of distributed multi-agent system design, as existing design methods are either suitable for small scale systematic synthesis, only dependent on large-scale topological models while oversimplifying the nodal dynamics, or fail to adapt to changing environments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1710072
Entities
People
- Panos Antsaklis
Organizations
- Army Contracting Command
- United States Army
- University of Notre Dame