Risk-Aware and Robust Decision Making in Multi-Agent MDPs

Abstract:

Autonomous systems increasingly operate alongside other agents, competing for shared resources or sharing control with human operators. In these settings, planning for the expected case is not enough, as rare but severe outcomes can have serious consequences. This talk presents research on multi-agent Markov decision processes (MDPs) with complex, real-world objectives. The first half of the talk addresses sharing scarce resources under uncertainty. I'll discuss several risk-aware objectives that balance safety with tractability, including chance constraints, Conditional Value-at-Risk (CVaR), and entropic risk. Further, I’ll present methods for synthesizing policies for these objectives in multi-agent MDPs.

The second half of the talk considers shared autonomy, where human and autonomous operators share control of semi-autonomous systems. I'll discuss how MDPs can be used to allocate control among operators, producing policies that are robust even when human operators are unknown or unreliable. Together, these results show how risk-awareness and robustness can be built into multi-agent planning without sacrificing tractability.

Read more about the Guest Speaker: Anna Gautier | Chalmers