Data Driven Decision Making Under Uncertainty With Applications In Healthcare And Energy Management

Author : Charalampos Andriotis
Publisher :
ISBN 13 :
Total Pages : pages
Book Rating : 4.:/5 (111 download)

DOWNLOAD NOW!

Download or read book Data Driven Decision Making Under Uncertainty for Intelligent Life-cycle Control of the Built Environment written by Charalampos Andriotis and published by . This book was released on 2019 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: This dissertation provides novel frameworks for data-driven probabilistic performance-based assessments and optimal or near-optimal stochastic control strategies for structural, infrastructural and other engineering systems. The goal of this research is to enable efficient and robust structural performance predictions and optimized decisions over the entire operating life of systems, by developing advanced statistical learning models, machine learning formulations and Artificial Intelligence (AI) algorithms, in order to contribute to a future of smart and sustainable infrastructure. To this end, the developed approaches build upon and extend well-established statistical modeling frameworks, infuse intelligence to structural informatics through newly introduced schemes for structural data mining and processing, provide comprehensive solutions to challenging life-cycle objectives, and support complex decisions in previously intractable sequential decision-making problems through novel AI-aided algorithms and theoretical concepts.Efficient assessment of various societal, environmental and economic losses necessitates adept statistical and learning models, able to consistently capture longitudinal dependencies in data and translate multivariate information in structural condition and performance metrics. This dissertation addresses this need, within a softmax regression fragility analysis framework that avoids fragility function crossing inconsistencies and scales well in high-dimensional intensity measure spaces with multiple structural states. Moreover, softmax-based fragility functions are generalized by advanced statistical learning and deep learning formulations that employ Dynamic Bayesian Networks (DBNs), in the form of Dependent Markov Models (DMMs) and Dependent Hidden Markov Models (DHMMs), as well as Recurrent Neural Network (RNN) architectures. The above considerably extend and generalize the framework of probabilistic performance engineering, with theoretically consistent multi-state multi-variate fragility functions, which also have multi-step predictive capabilities in time. The hidden spaces of DHMMs and RNNs are shown to be able to encode noisy input to noisy output sequences through structured hidden spaces. It turns out that the Markovian properties of these spaces can portray damage-consistent dynamics, whereas they are directly pertinent to the input required in advanced decision frameworks that employ Markovian processes for decision-making either under full, partial, or mixed observability assumptions.Hidden Markov models equipped with costs and control actions can provide a theoretically neat and computationally robust framework for sequential decision-making problems under uncertainty, through Partially Observable Markov Decision Processes (POMDPs). This research casts stochastic control problems for determination of optimal or near-optimal life-cycle maintenance and inspection strategies within the premises of POMDPs. Specialized formulations of full or mixed observability are also developed, through Markov Decision Processes (MDPs) or Mixed Observability Markov Decision Processes (MOMDPs), respectively. Along these lines, this research enables decision-support systems which can operate in stochastic engineering environments with uncertain action outcomes and noisy real-time observations, having global optimality guarantees as a result of the relevant underlying dynamic programming formulations introduced and, in many cases, well-defined performance bounds. In the same vein, the Value of Information (VoI) and the Value of Structural Health Monitoring (VoSHM) are quantified and a straightforward definition for the expected life-cycle gains of different observational and monitoring options is established and evaluated. Formulating VoI and VoSHM within the framework of POMDPs, the estimates of these metrics depict value gaps between the optimal life-cycle strategies of the examined options, thus also being able to provide bounds on the respective gains.For small- to medium-scale systems, solutions to the life-cycle optimization problems are derived by point-based solution schemes which provide efficient exploration heuristics, value function updates over the POMDP belief-space, vector compression techniques and convergence properties. For large-scale multi-component engineering systems that form large state and action spaces, such point-based schemes are however impractical as they require explicit prior information of the system dynamics model. To this end, the Deep Centralized Multi-agent Actor Critic (DCMAC) is developed herein and implemented in the solution procedure. DCMAC is an efficient off-policy actor-critic Deep Reinforcement Learning (DRL) algorithm with experience replay. DCMAC alleviates the curse of dimensionality related to state, observation and actions spaces of multi-component systems through deep network approximators and a factorized representation of the actor. DCMAC interacts directly with the simulator, thus avoiding the need for full and explicit model-based knowledge of the system dynamics, and operates in the POMDP belief space, by encoding sequences of actions and observations in belief vectors through Bayesian updates. Overall, DCMAC is able to efficiently tackle the state and action space scalability issues, as well as the potential model unavailability at the system level, all of which often make the decision problems of large multi-component systems hard to solve, if not intractable, by conventional machine learning schemes and other life-cycle optimization methodologies.All developed methods and frameworks are rigorously evaluated in relevant numerical applications and their strengths, limitations and broader capabilities are highlighted and discussed. Results demonstrate the effectiveness of the proposed models, solution procedures and algorithmic schemes, in enabling efficient data-driven probabilistic predictions and structural informatics, as well as comprehensive optimal or near-optimal stochastic control strategies for engineering systems. Overall, the originally developed statistical and machine learning models, in conjunction with the dedicated AI-aided algorithms, can ensure advanced and sophisticated solutions and open numerous new scientific paths towards smart cities, intelligent infrastructure, and autonomous control of the built environment.