Extend robust counterfactual interventions to complex temporal-logic properties

Develop an extension of the robust counterfactual intervention framework for uncertain Markov decision processes that handles complex temporal-logic properties by encoding them as reachability objectives through the product of a counterfactual interval Markov decision process and a reward machine.

Background

The paper formulates minimal robust counterfactual interventions primarily for threshold constraints on expected cumulative reward. It notes that reach-avoid properties can also be represented through an appropriate Bellman recursion, while more complex temporal-logic specifications require a product construction with a reward machine. The general treatment of such properties is not developed within the paper and is explicitly deferred to future work, leaving open the extension of the proposed intervention-computation procedure beyond expected-return objectives.

References

Properties like reach-avoid can be captured this way via a suitable Bellman recursion, more complex temporal logic properties can be encoded as reachability by taking the product of the CFIMDP with a reward machine citep{icarte2022reward}, which we leave for future work.

— Robust Counterfactuals for Markov Decision Processes under Statistical and Causal Uncertainties  (2610.20210 - Lally et al., 14 Oct 2026) in Section 4.3, “Minimal Robust Counterfactual Interventions”