Belabbas to research multi-agent systems with localized objectives
Belabbas is working to understand how the local interactions between agents, or any entities working within a system, can shape a desired global behavior and to discover how much information about this global behavior individual agents need to cooperate successfully. For example, people in groups or societies have their own life objectives, goals and behaviors, but they’re all also working toward larger global objectives such as keeping the economy running or taking care of the environment. Belabbas is asking how these two mesh together.
“We know the capability of individuals is important, but what we want to understand is the role of the network and seeing if it’s making the best of their capabilities, restricting the individuals’ capabilities or if the network isn’t rich enough to make the best of what the individuals have,” Belabbas said.
Belabbas came upon this problem after studying the formations of agents, such as birds or unmanned aerial vehicles (UAVs), flying through the air. While he worked to create control laws that explained the type of behavior encountered, such as creating a certain flying formation, he realized that some of the agents needed to know what their neighbors were trying to achieve in order to make the formation possible, yet they did not need to know anything else about these neighbors, not even their current positions. When looking at the theoretical side of control theory, he realized that the tools to analyze these systems were lacking.
The mathematical models behind systems where local interactions lead to global behavior are all very similar, whether it’s vehicles driving on an interstate, birds flying through the air or concentrations of certain proteins or enzymes in the body.
Belabbas, along with graduate student Artur Kirkoryan, are working to understand the mathematics behind these actions and to uncover fundamental laws that tell us how these interactions work.
“We want to look at localized objectives, which is something that has never been done before,” Belabbas said. “The idea is to understand how local objectives of the individual agents and the global objective of the group can go together. For example, is it possible to achieve the global objective without letting any of the agents know what it is?”
Most of the current work on multi-agent control focuses on issues relating to the agents having partial knowledge about what their neighbors are currently doing, but it is often implicitly assumed that everyone knows what the group wants to do. There is now a need to develop theories focusing on the objective of the ensemble and how it relates to the objectives of the agents.
“The novelty of the point of view adopted here lies in the handling of information about the system,” Belabbas said. “Beyond the usual restrictions on information available to the agents about the state of other agents, we integrate in the analysis restrictions on information about the global objective of the system.”
While the immediate impact of a fundamental law describing these interactions is unknown, Belabbas sees his theories being applicable in areas such as traffic management or, more generally, any type of human organization.
“You always want to work on a problem that will have an impact down the road and a paradigm that is here to stay is the one of collaboration,” Belabbas said.
Belabbas is a researcher in the university's Coordinated Science Laboratory (CSL).