151: Dueling Deep Q-Networks  in Policy Evaluation: Reinforcement Learning for Dynamic Decision-Making

This poster examines the use of Dueling Deep Q-Networks (Dueling DQNs)—a reinforcement learning method—in policy evaluation to support adaptive, real-time decision-making in complex systems. Traditional evaluation models often rely on static or linear frameworks, which can limit their effectiveness in dynamic environments with feedback loops and uncertainty. Dueling DQNs overcome these limitations by separately estimating a state’s overall value and the advantage of specific actions, enabling more nuanced policy modeling.

By simulating policy environments and learning from outcomes, Dueling DQNs forecast long-term impacts and suggest context-sensitive interventions. A public health case study on resource allocation illustrates the model’s predictive power, adaptability, and usefulness for evaluators. This paper advances the dialogue between machine learning and evaluation, highlighting not just technical capabilities but also ethical imperatives like equity, transparency, and utility.