CPS: Small: Multi-Human Assisted Learning for Multi-Agent Systems using Intrinsically Generated Event-Related EEG Potentials

该项目的目标是允许人类在网络物理系统中协助机器学习算法。该方法涉及使用基于脑电(EEG)的人在环路中的脑波来为学习算法生成反馈。机器学习解决方案对于监控、检测和优化复杂的网络物理系统(CP)特别有用。然而，一些跨越自然语言处理、自动语言翻译和理解场景或图像中的内容等领域的重要问题仍然超出了真正的机器学习算法的范围，需要人类参与。该项目涉及与佐治亚理工学院的Create-X和Venture Lab计划合作，将拟议研究的成果商业化。本课题考虑利用人脑电信号中的错误相关负性(ERN)来本质上识别机器的错误行为，并将其作为机器强化学习(RL)算法的奖励函数，以提高机器的智能性。这项研究的另一个新方面是考虑多个人在回路中提供反馈，这些反馈被嵌入在CPS中的学习算法用作奖励函数。最后，该项目还在探索收敛时间加速和深度学习技术的好处，以从EEG中提取独立于典型ERN分量峰的特征。这项研究是在现实世界CP的背景下使用基于游戏代理的分析和实验分析进行评估的。该奖项反映了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 goal of this project is to allow humans to assist machine learning algorithms in a Cyber-Physical System. The approach involves the use of electroencephalogram (EEG) based brain waves of the human-in-the-loop to generate feedback for the learning algorithms. Machine learning solutions are particularly useful for the monitoring, instrumenting, and optimization of complex cyber physical systems (CPS). However, several important problems spanning domains such as natural language processing, automated language translation, and understanding what is in a scene or image remain beyond the scope of true machine learning algorithms, and require human participation. The project involves collaboration with Georgia Tech's Create-X and Venture Lab programs to commercialize the outcomes of the proposed research. This project considers recognizing the erroneous behavior of the machine intrinsically by the error related negativity (ERN) in the human EEG signals, which are then used as the reward function for the reinforcement learning (RL) algorithm of the machine to improve its intelligence. Another novel aspect of the research is the consideration of multiple humans-in-the-loop providing feedback that is used as a reward function by the learning algorithms embedded within the CPS. Finally, the project also is exploring the benefits of convergence time acceleration and deep learning techniques to extract features from EEG that are independent of canonical ERN component peaks. The research is evaluated using both a game proxy based analysis and experimental analysis within the context of real-world CPS.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.