Planning under Distribution Shifts with Causal POMDPs

Researchers Matteo Ceriscioli and Karthika Mohan have introduced a novel theoretical framework, detailed in their paper 'Planning under Distribution Shifts with Causal POMDPs,' which tackles a core challenge in AI: maintaining robust planning when the real world changes. The work, set to appear at the 36th International Conference on Automated Planning and Scheduling (ICAPS-26), addresses the problem where AI agents' learned strategies fail because the environment's state distribution or dynamics shift. Their solution integrates causal knowledge into Partially Observable Markov Decision Processes (POMDPs), allowing shifts to be formally represented as interventions on a causal graph. This enables the system to actively hypothesize about which environmental components have changed and evaluate plans under these new conditions.

The technical breakthrough lies in proving that the value function in this augmented 'Causal POMDP' framework remains piecewise linear and convex (PWLC) in the belief space. This preservation is critical because it means the complex problem of planning under uncertainty and change retains mathematical tractability. Existing, efficient POMDP solvers that rely on α-vector representations can thus be adapted to work within this new framework without becoming computationally intractable. The approach maintains and updates a joint belief over both the latent state of the world and the identity of the underlying domain (i.e., which shift has occurred). This work provides a formal bridge between causal reasoning and sequential decision-making, offering a principled path toward creating AI agents that can adapt their plans when faced with unforeseen changes, a necessity for reliable deployment in dynamic real-world scenarios.