Inference in multi-agent systems


Multi-agent systems are systems where typically a number of agents perform certain task(s) with or without a central controller. Naturally a large number of real world systems qualify to be multi-agent systems. For example, swarms (robot or animal), games (cooperative or non-cooperative), sensor networks et cetera. The agents in a multi-agent system influence each other through an appropriate dependence structure, often described as a function of individual agent’s capacity to interact with others (directly or indirectly) or neighborhood relations.

From an engineering point of view, designing a multi-agent system so as to achieve a certain task is important and is studied across applications of multi-agent systems. However, the prime interest of this project lies in the inferential aspects of multi-agent systems in general. As a step towards this objective, we begin by modeling the multi-agent system as a (multi-agent) Markov Decision Process (MDP) or a Partially Observable Markov Decision Process (POMDP) whichever is suitable.


The main objective of this project is to find methods of inference for this setup. Typically only (noisy) data will be at our disposal to infer from. Parameters of interest are often, but not limited to, the associated policy or the reward function. We are also interested to study asymptotic behavior of such a multi-agent system. The limiting behavior when the number is agents is increasingly large is also of importance and is one of the objectives of this project. The project will involve writing computer programs, preferably in Matlab and also developing inference algorithms.


Students are expected to possess good programming skills. Also, willingness to learn different topics in disparate fields, as and when required, is solicited.


Wasiur Rahman Khuda Bukhsh