New Perspectives on Multiagent Dynamics
Abstract
This project fo. for research on network-based multiagent dynamics. Our ambition is to uncover basic principles that underlie the emergence of collective behavior from local interaction among autonomous agents. To carry out this ``local-to-global" research program, we plan to draw on techniques from dynamical systems, statistical mechanics, and algorithms. We intend to target a number of specific applications in opinion dynamics, consensus systems, swarming, iterated learning, and time-varying social networks. This list might suggest a project that divides neatly into topics sharing only a vague family resemblance. This is not the case. The emphasis of our work is on methods that cut horizontally across a diversity of application areas. We believe that the weak link in the field of multiagent dynamics is the current dearth of general techniques. Most systems are investigated in an ad hoc fashion. Our guiding objective, therefore, is to build a suite of analytical tools applicable to a broad spectrum of systems. The project consists of two overlapping parts and an applications component. egin{itemize} item[1.] {sc Opinion dynamics}: How do groups reach consensus without a leader? Under what conditions do they polarize into separate clusters? Most models of opinion dynamics treat agents as points in a metric space that move about under the influence of neighboring agents. This gives us an ideal platform to study network-based dynamics. Our attention will focus on two building blocks: the {em $s$-energy} and the concept of an {em algorithmic proof}. The former is a new type of generating function associated with dynamic networks (interestingly, it can also be used to rederive mixing bounds for random walks). Algorithmic proofs are protocols for distributed Lyapunov functions. Their investigation can be seen as an effort to build a bridge between the fields of dynamics and data structures. item[2.] {sc Influence systems}: As one of the most expressive models of multiagent dynamics, {em influence systems} are the perfect vehicle to investigate the conflict between energy and entropy that drives the bifurcation analysis of many multiagent systems. (Briefly, these seek to model open systems that absorb free energy into work while dissipating heat and producing entropy in the process). A significant part of this effort will be devoted to {em network-based renormalization}, a fundamental tool for reducing the dimension of large systems. This in turn will require further work on {em network parsing}, a new technique we have been developing for the hierarchical clustering of dynamic networks. item[3.] {sc Applications}: The $s$-energy has been found to play a key role in elucidating old questions about {em bird flocking}. Our interest in this area extends beyond bird behavior: indeed, swarming raises deep questions about non-Markovian averaging systems that might lead to useful, general insights into a wide variety of multiagent dynamical systems. Another application is in the area of {em iterated learning}. The motivation here is to understand the process of cultural transmission whereby a learner acquires a language from a teacher and, once done, proceeds to teach it to a third party, who in turn becomes the teacher, and so on. This process can be extended to arbitrary social networks~cite{kleinbergBook}. Under what conditions is iterated learning self-sustaining, meaning that the initial language is learned effectively by all the agents? end{itemize} We plan a parallel effort on all these questions on the expectation that progress in one area will inform our understanding in others. Computer simulations must be an integral part of the research effort but our focus is emphatically theoretical. Being aware that not everything interesting can be proven, we are comfortable switching back and forth between the theorem-proving work mode of the mathematician and the heuristic, phenomenologically-dr
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 14, 2019
- Source ID
- W911NF1710078XX0
Entities
People
- Bernard Chazelle
Organizations
- Army Contracting Command
- Defense Advanced Research Projects Agency
- Princeton University