EAGER: NSF2026: From Thinking to Inventing: Towards Creative Agents that Discover Novelty and Learn how to Accommodate it

EAGER：NSF2026：从思考到发明：走向发现新奇并学习如何适应它的创意代理

• 批准号: 2044786
• 负责人: Matthias Scheutz
• 金额: $ 30万
• 依托单位: Tufts University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-09-01 至 2023-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2044786&HistoricalAwards=false
• 关键词: EAGER; NSF2026; Thinking; Inventing; Towards

While recent successes demonstrate the potential of artificial intelligence (AI) technologies, there are tenacious challenges left to address before AI programs will be able to reach the cognitive flexibility and sophistication of the human mind. One of the most pressing problems is how to deal with the unknown, with novel contexts or aspects of the world for which an AI system was not originally designed. Such "open-world" AI is still in its infancy and existing AI techniques do not easily transfer to open worlds. The goal of this project is to investigate ways in which AI agents can be endowed with creative problem-solving skills that allow them to negotiate open worlds and invent new tools, concepts, and eventually theories. Such a system could be the next major disruptive technology in AI, enabling not only long-term autonomy and resilience of robots in the light of faults and unexpected events, but more importantly, providing technology that could accelerate the solution of pressing problems that humanity is currently able to solve only slowly, if at all.Current AI algorithms rely on having complete models of the task and domain in which they are supposed to operate. If information is missing, they are not able to plan how to acquire it and extend their knowledge. Yet, systems that can determine that they are missing critical knowledge could use that information to guide their knowledge acquisition process and possibly develop creative approaches for inventing new tools and theories. This project will develop an integrated problem-solving system that adapts and incorporates different learning techniques and deploys each of these methods in a targeted fashion based on the problem aspect the approach is best equipped to solve. This system will be extended with the ability to apply different strategies for experimenting (just like humans) with objects in its environment to discover their properties and the actions that can be performed with those objects to build up new knowledge about objects and their functions.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.

虽然最近的成功证明了人工智能（AI）技术的潜力，但在AI程序能够达到人类思维的认知灵活性和复杂性之前，还有一些棘手的挑战需要解决。 最紧迫的问题之一是如何处理未知的事物，以及人工智能系统最初设计时没有考虑的新环境或世界方面。 这种“开放世界”的人工智能仍处于起步阶段，现有的人工智能技术不容易转移到开放世界。 该项目的目标是研究如何赋予人工智能智能体创造性的解决问题的技能，使他们能够在开放的世界中进行谈判，并发明新的工具，概念和最终的理论。 这样的系统可能是人工智能领域的下一个重大颠覆性技术，不仅可以让机器人在故障和意外事件发生时实现长期自主和恢复能力，更重要的是，它提供的技术可以加速解决人类目前只能缓慢解决的紧迫问题。当前的人工智能算法依赖于拥有任务和领域的完整模型，它们应该在其中运行。 如果信息缺失，他们就无法计划如何获取信息和扩展知识。 然而，能够确定他们缺少关键知识的系统可以使用这些信息来指导他们的知识获取过程，并可能开发创造性的方法来发明新的工具和理论。 该项目将开发一个集成的问题解决系统，该系统适应并结合不同的学习技术，并根据该方法最适合解决的问题方面，以有针对性的方式部署这些方法中的每一种。 该系统将被扩展为应用不同策略进行实验的能力（就像人类一样）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进行评估，被认为值得支持的搜索.

项目成果

BIPLEX: Creative Problem-Solving by Planning for Experimentation

• 发表时间: 2022
• 作者: Vasanth Sarathy;Matthias Scheutz

SPOTTER: Extending Symbolic Planning Operators through Targeted Reinforcement Learning

SPOTTER：通过有针对性的强化学习扩展符号规划算子

• DOI: 10.5555/3463952.3464062
• 发表时间: 2021
• 期刊: AAMAS Conference proceedings
• 作者: Sarathy, Vasanth;Kasenberg, Daniel;Goel, Shivam;Sinapov, Jivko;Scheutz, Matthias
• 通讯作者: Scheutz, Matthias