Strategically-Linked Decisions in Long-Term Planning and Reinforcement Learning

ICLR 2026 Conference Submission14392 Authors

Keywords: strategic decision-making, long-term planning, reinforcement learning, explainability, safe policy improvement

Abstract: Long-term planning, as in reinforcement learning (RL), is often hard to interpret as it involves strategies: collections of actions that work toward a goal with potentially complex dependencies. In particular, some actions are taken at the expense of short-term benefit to enable future actions with even greater returns. In this paper, we quantify such dependencies between planned actions with *strategic link scores*: the drop in the likelihood of an earlier action under the constraint that a follow-up action is no longer available. We use strategic link scores to (i) explain black-box RL agents by identifying strategically-linked pairs among decisions they make, and (ii) improve the worst-case performance of decision support systems by distinguishing whether recommended actions can be adopted as standalone improvements, or whether they are strategically linked hence require a commitment to a broader strategy to be effective. We demonstrate these use cases with maze-solving and chess-playing examples as well as simulated healthcare and traffic environments.

Primary Area: reinforcement learning

Submission Number: 14392