Collaborative Research: CPS: Small: An Integrated Reactive and Proactive Adversarial Learning for Cyber-Physical-Human Systems

自动驾驶汽车的逐步部署将不可避免地导致一类新的重要的网络-物理-人类系统的出现，在这些系统中，自动驾驶车辆通过车载传感器或车与车之间的通信与人类驾驶的车辆互动。强化学习和控制理论可以帮助满足自动驾驶车辆的实时决策和5级自主的安全要求。然而，众所周知，传统的强化学习策略容易受到观测的对抗性或非对抗性扰动，类似于分类器和/或学习的奖励(包)丢弃的对抗性例子。随着使用开放的通信和控制平台实现自主变得至关重要，以及该行业继续投资于此类系统，人们对解决复原力的担忧加剧了这类问题。决策机制旨在结合敏捷性和强化学习的帮助，允许自我适应、自我修复和自我优化。这项研究将有助于和统一几个不同领域的知识体系，包括强化学习、安全、自动控制和运输，以实现人在回路中的弹性自主。在这个项目中，为了对抗动作和观察操纵以及奖赏下降，主要研究人员将利用主动切换策略，目标是(I)在闭环系统强化学习机制中为对抗性输入和奖赏下降提供稳健性，(Ii)通过欺骗增加操纵的成本，(Iii)限制易受攻击的动作和观察的暴露，以及(Iv)提供稳定性、最优性和稳健性保证。最终，研究人员将对上述每个领域做出基本贡献，并将这些领域合并，以提供一个独特的综合框架。该项目的成果将从伦理角度提高人们对自主技术的信心，为减少事故提供基础。拟议的框架可以扩展到全球经济的其他关键推动因素，包括智能和互联城市、医疗保健以及智能系统的网络行动，同时减少环境污染，最大限度地减少环境对人类健康的不利影响。该项目将通过良好协调、水平适当的参与研究和教育活动，培养来自不同水平、年龄和文化的下一代学生，同时为学生提供一个独特的机会，让他们欣赏高效、自主和低成本的设计。该项目还将有助于未来的工程课程，追求研究和教育的实质性整合，并提供机会让来自代表性不足群体的学生参与进来。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 gradual deployment of self-driving cars will inevitably lead to the emergence of a new important class of cyber-physical-human systems where autonomous vehicles interact with human-driven vehicles via on-board sensors or vehicle-to-vehicle communications. Reinforcement learning along with control theory can help meet the safety requirements for real-time decision making and Level 5 autonomy in self-driving vehicles. However, it is widely known that conventional reinforcement learning policies are vulnerable to adversarial or non-adversarial perturbations to their observations, similar to adversarial examples for classifiers and/or reward (packet) drops of the learning. Such issues are exacerbated by concerns of addressing resiliency as the use of open communication and control platforms for autonomy becomes essential, and as the industry continues to invest in such systems. Decision making mechanisms, designed to incorporate agility with the help of reinforcement learning, allow self-adaptation, self-healing, and self-optimization. This research will contribute and unify the body of knowledge of several diverse fields including reinforcement learning, security, automatic control, and transportation for resilient autonomy with humans-in-the-loop.In this project, to counter action and observation manipulation as well as reward drops, the principal investigators will leverage proactive switching policies that aim (i) to provide robustness to adversarial inputs and reward drops in the closed-loop reinforcement learning mechanisms, (ii) to increase the cost of manipulation by deception, (iii) to limit the exposure of vulnerable actions and observations, and (iv) to provide stability, optimality, and robustness guarantees. Ultimately, the investigators will develop fundamental contributions to each of the above-mentioned fields and amalgamate these fields to provide a unique synthesis framework. The outcomes of this project will increase levels of confidence in autonomous technologies from ethical perspectives by providing an underpinning for curtailing accidents. The proposed framework can be extended to other key enablers of the global economy, including smart and connected cities, healthcare, and networked actions of smart systems while decreasing environmental pollution and minimizing the adverse environmental impacts on human health. The project will train the next generation of students from various levels, ages, and cultures through well-coordinated, level appropriate involvement in research and educational activities while providing a unique opportunity for the students to appreciate efficient, autonomous, and low-cost designs. This project will also contribute to future engineering curricula, pursue a substantial integration of research and education, and provide opportunities to engage students from the underrepresented group.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.