Some Preliminary Studies on Controlling Agents in Multi-Agent Systems: From Constraint Satisfaction to Stochastic Games

PhD Thesis Proposal Defence


Title: "Some Preliminary Studies on Controlling Agents in Multi-Agent Systems:
From Constraint Satisfaction to Stochastic Games"

by

Mr. Fengming ZHU


Abstract:

Multi-agent and multi-robot systems have garnered increasing attention across
both theoretical domains, e.g., control theory and game theory, and practical
applications, e.g., warehouse automation and human-robot collaboration. This
preliminary study explores two complementary research directions toward the
problem of controlling those agents in such systems.

First, we investigate centralized control of robotic agents, framed as a
universal planning problem that can be interpreted through the lens of
constraint satisfaction. We therefore adopt logic programming as a
declarative approach to model and solve the problem. However, this
declarative paradigm suffers from the well-known grounding bottleneck, which
unavoidably leads to poor scalability. Informed by this limitation, we later
develop a rule-based but imperative system for warehouse automation in an
industrial scenario, where multi-robot planning in a real-world scale is even
interleaved with online task scheduling. It turns out to be beneficial if one
also takes the advantage of warehouse layouts while remaining the flexibility
for varying scales of robot fleets.

Secondly, we are interested in the problem of controlling one single agent
with the presence of multiple other agents about whom the protagonist agent
only has limited prior knowledge. By modelling this setting as a strategic
planning problem with an underlying skeleton of stochastic games, we manage
to draw a unified framework to encompass a spectrum of theoretic
formulations, such that a family of induced planners can be implemented and
evaluated in a common language. Importantly, our computational framework can
be carried over from stationary (0-memory) strategies to general K-memory
strategies. We also formally show that best responding to mixed K-memory
strategies is significantly harder than best responding to a single
(potentially randomized) K-memory strategy.

Notably, our work exhibits some conceptual overlap with research in cognitive
science, particularly the study of Theory of Mind (ToM). While we do have
some preliminary theoretical proposal, we aim to report more supporting
evidence from empirical results in the near future.


Date:                   Friday, 16 January 2026

Time:                   10:00am - 12:00noon

Venue:                  Room 3494
                        Lifts 25/26

Committee Members:      Prof. Fangzhen Lin (Supervisor)
                        Prof. Nevin Zhang (Chairperson)
                        Prof. Janet Hsiao