CAREER: Infrastructure Management under Model Uncertainty: Adaptive Sequential Learning and Decision Making

维护和运营相互关联的基础设施，包括运输和能源系统，对于为社会提供适当的安全和功能至关重要。在对老化基础设施的组件的当前状况、其演变以及维护行动的结果具有高度不确定性的情况下，必须对老化基础设施做出关键决策。随着传感器、机器人检查员、通信和计算变得更便宜、更可靠，它们的使用可以集成到基础设施管理中，以改善维护政策，降低不确定性、风险和总体管理成本。充分利用这一潜力需要改变我们如何建模和处理基础设施决策的范式，以及如何在统计推断和顺序优化中采用和扩展新的计算方法来管理基础设施系统。该CAREER奖项通过开发先进的模型和算法，利用传感器部署，概率推理和维护和操作决策之间的相互依赖性，最终提高维护和操作目的的决策能力并降低成本。该研究将与管理系统和开发软件的公司合作，在三个案例研究中测试开发的框架，包括风电场，天然气管道网络和桥梁网络。支撑项目研究的同一框架也构成了基于主动情境学习活动的基础设施管理教学方法的基础，其中将开发游戏，让K-12和大学生参与互动环境中的虚拟基础设施管理者角色。PI将与卡内基梅隆大学的SEE（女孩暑期工程体验）计划密切合作，以激发对风险分析和基础设施系统决策的兴趣。 该项目的目标是开发可扩展的算法，用于在持续的模型不确定性下进行长期决策。这些算法的目标不仅是基础设施系统的运营和维护决策，还包括收集更多信息：放置传感器、安排检查、规划测试以及试验新的运营和传感技术。通过学习和行动的连续迭代，管理者可以自适应地优化资源配置。当关注大型互联系统时，优化变得更加复杂，但回报更高：当众多基础设施组件在统计上相互依赖时，由于空间接近，共同的压力源或共同的动态模型，信息可以在它们之间传播。这使得在一个位置收集的观测结果也可以用于推断其他位置的条件，更重要的是，更高级别的系统演化模型。通过识别值得检查的位置和组件，代理可以顺序地调整系统的探索，避免昂贵的过度仪表化，并引导管理层可持续地使用有限的资源。为了开发这些方法，本项目将利用贝叶斯分层建模、随机场建模、前后验和信息值分析、子模块分析、部分可观察马尔可夫决策过程的近似动态规划、在线优化和网络分析。

Maintenance and operation of interconnected infrastructures, including transportation and energy systems, are critical to provide adequate levels of safety and functionality to society. Critical decisions must be made about aging infrastructures under high uncertainty about the current condition of their components, their evolution, and the results of maintenance actions. As sensors, robotic inspectors, communication and computation become cheaper and more reliable, their use can be integrated into infrastructure management to improve maintenance policies and reduce uncertainty, risk and overall management cost. Full exploitation of this potential requires a change in paradigm in how we model and approach infrastructure decision making, and how novel computational methods may be adopted and extended in statistical inference and sequential optimization for managing infrastructure systems. This CAREER award exploits the interdependency among sensor deployment, probabilistic inference and decision making on maintenance and operation by developing advanced models and algorithms that would ultimately result in enhanced decision making for maintenance and operation purposes and reduced costs. The research will test the developed framework on three case studies including wind farms, networks of gas pipelines, and networks of bridges, in collaboration with companies that manage systems and develop software. The same framework underpinning the project research also forms the basis for a pedagogic approach on infrastructure management based on active, situated-learning activities, in which games will be developed to involve K-12 and college students with assigned roles as virtual infrastructure managers in an interactive environment. The PI will work closely with Carnegie Mellon's SEE (Summer Engineering Experience for Girls) program to inspire interest in risk analysis and infrastructure systems decision making. This project's goal is to develop scalable algorithms for long-term decision-making under persistent model uncertainty. These algorithms will target decisions about not only operation and maintenance of infrastructure systems, but also about the gathering of further information: placing sensors, scheduling inspections, planning tests and experimenting with new technologies for operation and sensing. By sequential iterations of learning and acting, managers can adaptively optimize resource allocation. Optimization becomes more complex, but with higher pay-off, when focusing on large interconnected systems: when numerous infrastructure components are statistically interdependent, because of spatial proximity, common stressors, or common dynamic models, information can propagate across them. This allows observations collected in one location to also be useful for inferring the conditions of others and, more importantly, the higher-level system evolution model. By identifying locations and components that deserve inspections, the agent can sequentially adapt the exploration of the system, avoid costly over-instrumentation, and guide management towards a sustainable use of the limited resources. To develop these approaches, this project will make use of Bayesian hierarchical modeling, random field modeling, pre-posterior and value-of-information analysis, sub-modularity analysis, approximate dynamic programming for Partially Observable Markov Decision Processes (POMDP), on-line optimization and network analysis.