CAREER: Adaptive Intelligent Agents for the Internet

该项目将有两个主要目标：(1)继续扩大对更先进的强化学习方案的开发和理解，这些方案已经证明了比传统的强化学习方案更能处理更大规模的问题和在大脑中模拟学习的各个方面的能力；(2)将先进的强化学习应用于开发更强大的互联网智能代理--并将这一应用作为开发更强大的设计的刺激。例如，现有的Web搜索引擎是一类极其重要的代理，但它们的能力也非常有限，因为它们倾向于依赖简单的字符串匹配技术，偶尔还会辅以-与-或运算符和有限类型的用户提供的线索。本工作将探索一种解决这一问题的新方法，该方法基于一个分布式的智能代理系统，该系统专注于不同的领域，但有效地学习相互理解。如果这一方法成功，它可能会产生进一步的影响，远远超出对更好的搜索的需要，因此，它最终可能有助于将因特网本身转变为一个大型分布式智能系统的愿景。该项目还将包括在教育中应用这些新兴工具的大量努力，以及与这项工作相关的新课程的开发。

This project will have two primary objectives: (1) to continue to extend the development and understanding of more advanced reinforcement learning schemes, which have demonstrated the ability to handle larger-scale problems and to mimic aspects of learning in the brain more than the conventional reinforcement learning schemes; (2) to apply advanced reinforcement learning to the development of more powerful intelligent agents for the Internet--and to use this application as a stimulus to the development of more powerful designs. For example, existing search engines for the Web are an extremely important class of agent, but also very limited in their capability because they tend to rely on simple string-mating techniques, occasionally supplemented by -and+ and -or+ operators and limited types of user-supplied clues. This work will explore a novel approach to this problem, based on a distributed system of intelligent agents which focus on different domains but learn to understand each other, in effects. If this approach is successful, it could have further implications well beyond the need for better search as such Ultimately it could contribute to the vision of converting the Internet itself into one large distributed intelligent system. This project will also include a substantial effort to apply these emerging tools in education, and the development of new courses related to this work.