Multi-Robot Collaboration for Trustworthy Lifelong Learning

多机器人协作实现值得信赖的终身学习

• 批准号: 2774392
• 金额: --
• 依托单位: University of Southampton
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2022
• 资助国家: 英国
• 起止时间: 2022 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2774392
• 关键词: Multi; Robot; Collaboration; Trustworthy; Lifelong

Robots require lifelong learning to acquire, refine, and translate knowledge from new experiences. However, by making decisions based on information that is acquired continuously, as opposed to learning from large-scale datasets that are available a priori, lifelong learning raises significant challenges for the assurance of trustworthy AI: Homogeneous data acquired continually through insufficiently diverse experiences can lead to biased decisions (e.g., a robot with experience limited to households cannot straightforwardly translate its prior knowledge when exposed to, e.g., public spaces). To ensure that a robot's reasoning is based on evidence from heterogenous experiences, distributed learning, e.g., federated learning, allows multiple robots expose to diverse scenarios to collaboratively share their individual knowledge. While distributed learning enables robots to think 'outside of the box', a risk remains that homogenous subgroups of robots instil systematic bias in the fused experience of the heterogenous cohort. This project will analyse the source of bias in distributed learning, and will develop novel algorithms that mitigate bias during the aggregation of experiences from multiple robots, while preserving fairness and ensuring data privacy.

机器人需要终身学习，从新的经验中获取、提炼和转化知识。然而，与从先验的大规模数据集中学习不同，终身学习基于不断获取的信息做出决策，为确保人工智能的可信赖性提出了重大挑战：通过不充分多样化的经验不断获得的同质数据可能导致有偏见的决策（例如，经验仅限于家庭的机器人在暴露于公共空间时无法直接翻译其先前的知识）。为了确保机器人的推理是基于来自异质经验的证据，分布式学习，例如联邦学习，允许多个机器人暴露在不同的场景中，以协作共享他们的个人知识。虽然分布式学习使机器人能够“跳出框框”思考，但风险仍然存在，即同质的机器人子组在异质队列的融合经验中灌输系统偏见。该项目将分析分布式学习中偏见的来源，并将开发新的算法，以减轻来自多个机器人的经验汇总过程中的偏见，同时保持公平性并确保数据隐私。