Decisions made at a given stage of a process can constrain or enable future actions, thereby influencing long-term outcomes. In many scientific domains, such as precision medicine, public policy, and economics, the quality of an initial decision cannot be evaluated solely by its immediate effect, but rather by its consequences across an entire sequence of future decision points. For example, an initially modestly effective chemotherapy option may lead to improved long-term survival when followed by an appropriate salvage regimen. This motivates statistical methods that explicitly account for downstream interventions, evolving covariate processes, and future decision rules. Within the framework of dynamic treatment regimes and reinforcement learning, estimation of optimal sequential decisions requires modeling both immediate and future conditional gains or rewards. In this talk, we will discuss Q-learning as a statistical learning approach for estimating optimal dynamic treatment regimes. I will emphasize its interpretation, implementation, and theoretical properties, as well as its strengths and limitations relative to alternative methods. The goal is to illustrate how forward-looking statistical decision strategies can yield improved long-term outcomes.
