<ul><li>Long-term planning in reinforcement learning involves finding strategies that collectively work toward a goal rather than solely optimizing immediate outcomes.</li><li>Quantifying dependencies between planned actions in a strategy is done through strategic link scores, measuring the drop in the likelihood of one decision if a follow-up decision is no longer available.</li><li>The utility of strategic link scores is demonstrated in three practical applications: explaining black-box RL agents, improving decision support systems, and characterizing planning processes of non-RL agents through interventions.</li><li>For instance, one application involves analyzing a traffic simulator by observing the emergent routing behavior of drivers through road closures to measure the effective planning horizon.</li></ul>

Strategically Linked Decisions in Long-Term Planning and Reinforcement Learning

Discover more