Collaborative Research: III: Medium: Designing AI Systems with Steerable Long-Term Dynamics

用户通过他们的选择(例如，点击、购买)提供的反馈是容易用于培训自主信息检索和推荐系统的最常见的数据类型之一，并且在在线平台中被广泛使用。然而，基于选择数据的天真的培训系统可能只会改善短期参与度，而不会改善平台的长期可持续性以及平台用户、内容提供商和其他利益相关者的长期利益。在这个问题和利益冲突的复杂空间中，不太可能有一个单一而紧凑的算法解决方案本质上是公平或最佳的-出于同样的原因，我们的法律法规和税收政策充满了相当大的库。相反，该项目开发了一个新的算法框架，以表达同样适用于人工智能系统的类似详细政策。这一框架为决策者提供了战略干预，可以预测地引导平台的长期动态，使他们不仅在短期内优化参与，而且还反映了监管平台的任何治理系统设定的长期价值观。为了实现这一目标，该项目为人工智能平台引入了一个宏观抽象层，在该抽象层下，可以通过宏观干预(例如，曝光分配、新内容的推广政策、反歧视监管)来衡量和影响长期目标(例如，用户满意度、项目公平性、供应商池规模)。由于平台在微观层面发挥作用，该项目开发了新的搜索和推荐方法，以最佳方式将宏观层面的干预措施分解为一系列微观层面的干预措施(例如，排名)。关键的技术挑战在于弥合宏观干预措施(例如，数周)和微观干预措施(例如，个别请求)在时间尺度上的不匹配，这是使用机器学习、因果推理和控制理论解决的。这一提法提供了一个技术抽象层，降低了宏观层面上人工和自动决策的复杂性，使战略推理和行动成为可能。最后，由于任何层面的最佳行动都依赖于公正和准确的估计，该项目开发了新的估计器，以抵消反馈循环中的偏差。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The feedback that users provide through their choices (e.g., clicks, purchases) is one of the most common types of data readily available for training autonomous information retrieval and recommendation systems, and it is widely used in online platforms. However, naively training systems based on choice data may only improve short-term engagement, but not the long-term sustainability of the platform and the long-term benefits to its users, content providers, and other stakeholders. In this complex space of problems and competing interests, it is unlikely that there is a single and compact algorithmic solution that is inherently fair or optimal --- for the same reason that our legal codes and tax policies fill sizable libraries. Instead, the project develops a new algorithmic framework to express similarly detailed policies also for AI systems. This framework provides decision-makers with strategic interventions that predictably steer the long-term dynamics of a platform so that they not only optimize engagement in the short term but additionally reflect long-term values set by whatever system of governance oversees the platform. To achieve this goal, the project introduces a macroscopic layer of abstraction for AI platforms under which long-term objectives (e.g., user satisfaction, item fairness, supplier pool size) can be measured and influenced through macroscopic interventions (e.g., exposure allocation, promotion policies for new content, anti-discrimination regulation). Since platforms act at the microscopic level, the project develops new search and recommendation methods that optimally break macro-level interventions into a sequence of micro-level interventions (e.g., rankings). The crucial technical challenge lies in bridging the mismatch in time scales between macro-level interventions (e.g., weeks) and micro-level interventions (e.g., individual requests), which is addressed using machine learning, causal inference, and control theory. This formulation provides a technical layer of abstraction that reduces complexity for both human and automated decision-making at the macro level, enabling strategic reasoning and action. Finally, since optimal actions at any level rely on unbiased and accurate estimates, the project develops new estimators that counteract biases in feedback loops.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.