EAGER: Collaborative Research: An Unified Learnable Roadmap for Sequential Decision Making in Relational Domains

该项目旨在开发新的算法和数据结构，用于学习和规划环境中的对象之间的一组关系。关系表示以简洁且易于解释的表示方式捕获对象之间的交互。非常适合关系表示的领域的示例包括智能无人机协助士兵，供应链管理中的活动，社交网络中的通信和友谊连接，以及跟踪视频中的个人和活动。 然而，机器学习和规划领域的最新进展，例如所谓的“深度神经网络”，采用简单的“平面”表示，其中世界状态是一串无法解释的位。 该项目将使机器学习和规划方法更容易使用，更强大，通过推广它们，使它们显式地与关系模型和数据一起工作。 该提案产生的方法、理论和数据将以几种积极的方式影响科学界，并将通过适当的网站公开提供。这项研究将通过参考期刊和会议记录传播，并提供给研究人员。拟议算法的代码和新基准问题的描述也将公开提供。调查人员将根据该项目带来的挑战和调查结果组织讲习班和辅导班。已经开发了许多特殊用途的解决方案来解决这些问题的一小部分，但没有通用工具可以利用机器学习的最新进展来解决这一系列问题。该提案旨在开发此类工具，利用研究人员先前在学习关系回归树方面的经验和强化学习的值函数近似经验。此外，该项目旨在为深度学习的最新进展（通常与关系表示不兼容）和关系学习的最新进展搭建桥梁。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project seeks to develop new algorithms and data structures for learning and planning in situations where the environment is represented with a set of relations between objects. Relational representations capture interactions between objects in a succinct and easily interpretable representation. Examples of domains that are well-suited to relational representations includes intelligent drones assisting soldiers, activities in a supply chain management, communication and friendship connections in a social network, and tracking individuals and activities in video. Most recent advances in machine learning and planning, such as so-called "deep neural networks", however, employ simple "flat" representations, where the state of the world is an uninterpreted string of bits. This project will make machine learning and planning methods easier to use and more robust by generalizing them so that they explicitly work with relational models and data. The methods, theory, and data resulting from this proposal will impact the scientific community in several positive ways and will be made publicly available through an appropriate website. The research will be disseminated through refereed journals and conference proceedings and made available to researchers. Code for the proposed algorithms and descriptions of new benchmark problems will also be made publicly available. The investigators will work on organizing workshops and tutorials based on the challenges and findings arising from this project. Many special purpose solutions have been developed to address small parts of these problems, but there are no general purpose tools that harness recent advances in machine learning to tackle this family of problems. This proposal seeks to develop such tools, drawing upon the investigators' prior experience in learning relational regression trees and experience in value function approximation for reinforcement learning. In addition, this project seeks to build a bridge between recent advances in deep learning, which generally has not been compatible with relational representations, and recent advances in relational learning.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.