Scholarly Works, Industrial and Systems Engineering

Permanent URI for this collection

Research articles, presentations, and other scholarship

Browse

Recent Submissions

Now showing 1 - 20 of 472
  • Trust at Your Own Peril: A Mixed Methods Exploration of the Ability of Large Language Models to Generate Expert-Like Systems Engineering Artifacts and a Characterization of Failure Modes
    Topcu, Taylan G.; Husain, Mohammed; Ofsa, Max; Wach, Paul (Wiley, 2025-02-21)
    Multi-purpose large language models (LLMs), a subset of generative artificial intelligence (AI), have recently made significant progress. While expectations for LLMs to assist systems engineering (SE) tasks are paramount; the interdisciplinary and complex nature of systems, along with the need to synthesize deep-domain knowledge and operational context, raise questions regarding the efficacy of LLMs to generate SE artifacts, particularly given that they are trained using data that is broadly available on the internet. To that end, we present results from an empirical exploration, where a human expert-generated SE artifact was taken as a benchmark, parsed, and fed into various LLMs through prompt engineering to generate segments of typical SE artifacts. This procedure was applied without any fine-tuning or calibration to document baseline LLM performance. We then adopted a two-fold mixed-methods approach to compare AI generated artifacts against the benchmark. First, we quantitatively compare the artifacts using natural language processing algorithms and find that when prompted carefully, the state-of-the-art algorithms cannot differentiate AI-generated artifacts from the human-expert benchmark. Second, we conduct a qualitative deep dive to investigate how they differ in terms of quality. We document that while the two-material appear very similar, AI generated artifacts exhibit serious failure modes that could be difficult to detect. We characterize these as: premature requirements definition, unsubstantiated numerical estimates, and propensity to overspecify. We contend that this study tells a cautionary tale about why the SE community must be more cautious adopting AI suggested feedback, at least when generated by multi-purpose LLMs.
  • Effect of processing parameters and thermal history on microstructure evolution and functional properties in laser powder bed fusion of 316L
    Deshmukh, Kaustubh; Riensche, Alex; Bevans, Ben; Lane, Ryan J.; Snyder, Kyle; Halliday, Harold (Scott); Williams, Christopher B.; Mirzaeifar, Reza; Rao, Prahalada (Elsevier, 2024-07-03)
    In this paper, we explain and quantify the causal effect of processing parameters and part-scale thermal history on the evolution of microstructure and mechanical properties in the laser powder bed fusion additive manufacturing of Stainless Steel 316L components. While previous works have correlated the processing parameters to flaw formation, microstructures evolved, and properties, a missing link is the understanding of the effect of thermal history. Accordingly, tensile test coupons were manufactured under varying processing conditions, and their microstructure-related attributes, e.g., grain morphology, size and texture; porosity; and microhardness were characterized. Additionally, the yield and tensile strengths of the samples were measured using digital image correlation. An experimentally validated computational model was used to predict the thermal history of each sample. The temperature gradients and sub-surface cooling rates ascertained from the model predictions were correlated with the experimentally characterized microstructure and mechanical properties. By elucidating the fundamental process-thermal-structure–property relationship, this work establishes the foundation for future physics-based prediction of microstructure and functional properties in laser powder bed fusion.
  • Sub-surface thermal measurement in additive manufacturing via machine learning-enabled high-resolution fiber optic sensing
    Wang, Rongxuan; Wang, Ruixuan; Dou, Chaoran; Yang, Shuo; Gnanasambandam, Raghav; Wang, Anbo; Kong, Zhenyu (James) (Springer Nature, 2024-08-31)
    Microstructures of additively manufactured metal parts are crucial since they determine the mechanical properties. The evolution of the microstructures during layer-wise printing is complex due to continuous re-melting and reheating effects. The current approach to studying this phenomenon relies on time-consuming numerical models such as finite element analysis due to the lack of effective sub-surface temperature measurement techniques. Attributed to the miniature footprint, chirped-fiber Bragg grating, a unique type of fiber optical sensor, has great potential to achieve this goal. However, using the traditional demodulationmethods, its spatial resolution is limited to the millimeter level. In addition, embedding it during laser additive manufacturing is challenging since the sensor is fragile. This paper implements a machine learning-assisted approach to demodulate the optical signal to thermal distribution and significantly improve spatial resolution to 28.8 μm from the original millimeter level. A sensor embedding technique is also developed to minimize damage to the sensor and part while ensuring close contact. The case study demonstrates the excellent performance of the proposed sensor in measuring sharp thermal gradients and fast cooling rates during the laser powder bed fusion. The developed sensor has a promising potential to study the fundamental physics of metal additive manufacturing processes.
  • An expandable machine learning-optimization framework to sequential decision-making
    Yilmaz, Dogacan; Büyüktahtakın, İ. Esra (Elsevier, 2024-04)
    We present an integrated prediction-optimization (PredOpt) framework to efficiently solve sequential decision-making problems by predicting the values of binary decision variables in an optimal solution. We address the key issues of sequential dependence, infeasibility, and generalization in machine learning (ML) to make predictions for optimal solutions to combinatorial problems. The sequential nature of the combinatorial optimization problems considered is captured with recurrent neural networks and a sliding-attention window. We integrate an attention-based encoder–decoder neural network architecture with an infeasibility-elimination and generalization framework to learn high-quality feasible solutions to time-dependent optimization problems. In this framework, the required level of predictions is optimized to eliminate the infeasibility of the ML predictions. These predictions are then fixed in mixed-integer programming (MIP) problems to solve them quickly with the aid of a commercial solver. We demonstrate our approach to tackling the two well-known dynamic NP-Hard optimization problems: multi-item capacitated lot-sizing (MCLSP) and multi-dimensional knapsack (MSMK). Our results show that models trained on shorter and smaller-dimensional instances can be successfully used to predict longer and larger-dimensional problems. The solution time can be reduced by three orders of magnitude with an average optimality gap below 0.1%. We compare PredOpt with various specially designed heuristics and show that our framework outperforms them. PredOpt can be advantageous for solving dynamic MIP problems that need to be solved instantly and repetitively.
  • Risk-averse multi-stage stochastic optimization for surveillance and operations planning of a forest insect infestation
    Bushaj, Sabah; Büyüktahtakın, İ. Esra; Haight, Robert G. (Elsevier, 2022-06)
    We derive a nested risk measure for a maximization problem and implement it in a scenario-based formulation of a multi-stage stochastic mixed-integer programming problem. We apply the risk-averse formulation to the surveillance and control of a non-native forest insect, the emerald ash borer (EAB), a wood-boring beetle native to Asia and recently discovered in North America. Spreading across the eastern United States and Canada, EAB has killed millions of ash trees and cost homeowners and local governments billions of dollars. We present a mean-Conditional Value-at-Risk (CVaR), multi-stage, stochastic mixed-integer programming model to optimize a manager’s decisions about surveillance and control of EAB. The objective is to maximize the benefits of healthy ash trees in forests and urban environments under a fixed budget. Combining the risk-neutral objective with a risk measure allows for a trade-off between the weighted expected benefits from ash trees and the expected risks associated with experiencing extremely damaging scenarios. We define scenario dominance cuts (sdc) for the maximization problem and under the decision-dependent uncertainty. We then solve the model using the sdc cutting plane algorithm for varying risk parameters. Computational results demonstrate that scenario dominance cuts significantly improve the solution performance relative to that of CPLEX. Our CVaR risk-averse approach also raises the objective value of the least-benefit scenarios compared to the risk-neutral model. Results show a shift in the optimal strategy from applying less expensive insecticide treatment to more costly tree removal as the manager becomes more risk-averse. We also find that risk-averse managers survey more often to reduce the risk of experiencing adverse outcomes.
  • A deep reinforcement learning framework for solving two-stage stochastic programs
    Yilmaz, Dogacan; Büyüktahtakın, İ. Esra (Springer, 2023-05-31)
    In this study, we present a deep reinforcement learning framework for solving scenario-based two-stage stochastic programming problems. Stochastic programs have numerous real-time applications, such as scheduling, disaster management, and route planning, yet they are computationally challenging to solve and require specially designed solution strategies such as hand-crafted heuristics. To the extent of our knowledge, this is the first study that decomposes two-stage stochastic programs with a multi-agent structure in a deep reinforcement learning algorithmic framework to solve them faster. Specifically, we propose a general two-stage deep reinforcement learning framework that can generate high-quality solutions within a fraction of a second, in which two different learning agents sequentially learn to solve each stage of the problem. The first-stage agent is trained with the feedback of the second-stage agent using a new policy gradient formulation since the decisions are interconnected through the stages. We demonstrate our framework through a general multi-dimensional stochastic knapsack problem. The results show that solution time can be reduced up to five orders of magnitude with sufficiently good optimality gaps of around 7%. Also, a decision-making agent can be trained with a few scenarios and can solve problems with many scenarios and achieve a significant reduction in solution times. Considering the vast state and action space of the problem of interest, the results show a promising direction for generating fast solutions for stochastic online optimization problems without expert knowledge.
  • COVID-19: Data-Driven optimal allocation of ventilator supply under uncertainty and risk
    Yin, Xuecheng; Büyüktahtakın, İ. Esra; Patel, Bhumi P. (Elsevier, 2021-12-01)
    This study presents a new risk-averse multi-stage stochastic epidemics-ventilator-logistics compartmental model to address the resource allocation challenges of mitigating COVID-19. This epidemiological logistics model involves the uncertainty of untested asymptomatic infections and incorporates short-term human migration. Disease transmission is also forecasted through a new formulation of transmission rates that evolve over space and time with respect to various non-pharmaceutical interventions, such as wearing masks, social distancing, and lockdown. The proposed multi-stage stochastic model overviews different scenarios on the number of asymptomatic individuals while optimizing the distribution of resources, such as ventilators, to minimize the total expected number of newly infected and deceased people. The Conditional Value at Risk (CVaR) is also incorporated into the multi-stage mean-risk model to allow for a trade-off between the weighted expected loss due to the outbreak and the expected risks associated with experiencing disastrous pandemic scenarios. We apply our multi-stage mean-risk epidemics-ventilator-logistics model to the case of controlling COVID-19 in highly-impacted counties of New York and New Jersey. We calibrate, validate, and test our model using actual infection, population, and migration data. We also define a new region-based sub-problem and bounds on the problem and then show their computational benefits in terms of the optimality and relaxation gaps. The computational results indicate that short-term migration influences the transmission of the disease significantly. The optimal number of ventilators allocated to each region depends on various factors, including the number of initial infections, disease transmission rates, initial ICU capacity, the population of a geographical location, and the availability of ventilator supply. Our data-driven modeling framework can be adapted to study the disease transmission dynamics and logistics of other similar epidemics and pandemics.
  • A multi-stage stochastic programming approach to epidemic resource allocation with equity considerations
    Yin, Xuecheng; Büyüktahtakın, İ. Esra (Springer, 2021-05-10)
    Existing compartmental models in epidemiology are limited in terms of optimizing the resource allocation to control an epidemic outbreak under disease growth uncertainty. In this study, we address this core limitation by presenting a multi-stage stochastic programming compartmental model, which integrates the uncertain disease progression and resource allocation to control an infectious disease outbreak. The proposed multi-stage stochastic program involves various disease growth scenarios and optimizes the distribution of treatment centers and resources while minimizing the total expected number of new infections and funerals. We define two new equity metrics, namely infection and capacity equity, and explicitly consider equity for allocating treatment funds and facilities over multiple time stages. We also study the multi-stage value of the stochastic solution (VSS), which demonstrates the superiority of the proposed stochastic programming model over its deterministic counterpart. We apply the proposed formulation to control the Ebola Virus Disease (EVD) in Guinea, Sierra Leone, and Liberia of West Africa to determine the optimal and fair resource-allocation strategies. Our model balances the proportion of infections over all regions, even without including the infection equity or prevalence equity constraints. Model results also show that allocating treatment resources proportional to population is sub-optimal, and enforcing such a resource allocation policy might adversely impact the total number of infections and deaths, and thus resulting in a high cost that we have to pay for the fairness. Our multi-stage stochastic epidemic-logistics model is practical and can be adapted to control other infectious diseases in meta-populations and dynamically evolving situations.
  • Risk-averse multi-stage stochastic programming to optimizing vaccine allocation and treatment logistics for effective epidemic response
    Yin, Xuecheng; Büyüktahtakın, İ. Esra (Taylor & Francis, 2021-07-24)
    Existing compartmental-logistics models in epidemics control are limited in terms of optimizing the allocation of vaccines and treatment resources under a risk-averse objective. In this paper, we present a data-driven, mean-risk, multi-stage, stochastic epidemics-vaccination-logistics model that evaluates various disease growth scenarios under the Conditional Value-at-Risk (CVaR) risk measure to optimize the distribution of treatment centers, resources, and vaccines, while minimizing the total expected number of infections, deaths, and close contacts of infected people under a limited budget. We integrate a new ring vaccination compartment into a Susceptible-Infected-Treated-Recovered-Funeral-Burial epidemics-logistics model. Our formulation involves uncertainty both in the vaccine supply and the disease transmission rate. Here, we also consider the risk of experiencing scenarios that lead to adverse outcomes in terms of the number of infected and dead people due to the epidemic. Combining the risk-neutral objective with a risk measure allows for a tradeoff between the weighted expected impact of the outbreak and the expected risks associated with experiencing extremely disastrous scenarios. We incorporate human mobility into the model and develop a new method to estimate the migration rate between each region when data on migration rates is not available. We apply our multi-stage stochastic mixed-integer programming model to the case of controlling the 2018–2020 Ebola Virus Disease (EVD) in the Democratic Republic of the Congo (DRC) using real data. Our results show that increasing the risk-aversion by emphasizing potentially disastrous outbreak scenarios reduces the expected risk related to adverse scenarios at the price of the increased expected number of infections and deaths over all possible scenarios. We also find that isolating and treating infected individuals are the most efficient ways to slow the transmission of the disease, while vaccination is supplementary to primary interventions on reducing the number of infections. Furthermore, our analysis indicates that vaccine acceptance rates affect the optimal vaccine allocation only at the initial stages of the vaccine rollout under a tight vaccine supply.
  • A game-theoretic approach to incentivize landowners to mitigate an emerald ash borer outbreak
    Chen, Chen; Cai, Wenbo; Büyüktahtakın, İ. Esra; Haight, Robert G. (Taylor & Francis, 2023-09-13)
    This article addresses the challenge posed by the Emerald Ash Borer (EAB), a wood–boring insect that threatens to kill ash trees, one of the North America’s most vital tree genera. Current strategies include monitoring, treatment, and removal. However, the absence of a private-public partnership hinders progress on private ash trees. We propose two cost-sharing programs where local governments reimburse landowners for their management costs. This approach considers the EAB’s dynamic growth over two periods based on different treatment decisions. Two mathematical models are developed for designing reimbursements: one based on the number of infested trees and another on the number of treated trees. We derive analytical solutions for the optimal treatment decisions and reimbursements in the first period. Our study reveals that treatment effectiveness and the likelihood of new infestations in the second period influence the optimal decisions. Comparing the reimbursement models, the treatment-based program proves more effective, encouraging landowners to treat more trees with higher reimbursements and overall benefits. Further, we show that continuing EAB treatment beyond the 2-year cost-sharing program is expected to yield superior long-term benefits. The approach seeks to foster private-public partnerships in addressing environmental challenges through resource sharing, such as managing water, land, and wildfires.
  • COVID-19: Agent-based simulation-optimization to vaccine center location vaccine allocation problem
    Yin, Xuecheng; Bushaj, Sabah; Yuan, Yue; Büyüktahtakın, İ. Esra (Taylor & Francis, 2023-08-10)
    This article presents an agent-based simulation-optimization modeling and algorithmic framework to determine the optimal vaccine center location and vaccine allocation strategies under budget constraints during an epidemic outbreak. Both simulation and optimization models incorporate population health dynamics, such as susceptible (S), vaccinated (V), infected (I) and recovered (R), while their integrated utilization focuses on the COVID-19 vaccine allocation challenges. We first formulate a dynamic location–allocation Mixed-Integer Programming (MIP) model, which determines the optimal vaccination center locations and vaccines allocated to vaccination centers, pharmacies, and health centers in a multi-period setting in each region over a geographical location. We then extend the agent-based epidemiological simulation model of COVID-19 (Covasim) by adding new vaccination compartments representing people who take the first vaccine shot and the first two shots. The Covasim involves complex disease transmission contact networks, including households, schools, and workplaces, and demographics, such as age-based disease transmission parameters. We combine the extended Covasim with the vaccination center location-allocation MIP model into one single simulation-optimization framework, which works iteratively forward and backward in time to determine the optimal vaccine allocation under varying disease dynamics. The agent-based simulation captures the inherent uncertainty in disease progression and forecasts the refined number of susceptible individuals and infections for the current time period to be used as an input into the optimization. We calibrate, validate, and test our simulation-optimization vaccine allocation model using the COVID-19 data and vaccine distribution case study in New Jersey. The resulting insights support ongoing mass vaccination efforts to mitigate the impact of the pandemic on public health, while the simulation-optimization algorithmic framework could be generalized for other epidemics.
  • A Multistage Stochastic Programming Approach to the Optimal Surveillance and Control of the Emerald Ash Borer in Cities
    Kibis, Eyyub Y.; Büyüktahtakın, İ. Esra; Haight, Robert G.; Akhundov, Najmaddin; Knight, Kathleen; Flower, Charles E. (INFORMS, 2020-10-12)
    Emerald ash borer (EAB), a wood-boring insect native to Asia and invading North America, has killed untold millions of high-value ash trees that shade streets, homes, and parks and caused significant economic damage in cities of the United States. Local actions to reduce damage include surveillance to find EAB and control to slow its spread. We present a multistage stochastic mixed-integer programming (M-SMIP) model for the optimization of surveillance, treatment, and removal of ash trees in cities. Decision-dependent uncertainty is modeled by representing surveillance decisions and the realizations of the uncertain infestation parameter contingent on surveillance as branches in the M-SMIP scenario tree. The objective is to allocate resources to surveillance and control over space and time to maximize public benefits. We develop a new cutting-plane algorithm to strengthen the M-SMIP formulation and facilitate an optimal solution. We calibrate and validate our model of ash dynamics using seven years of observational data and apply the optimization model to a possible infestation in Burnsville, Minnesota. Proposed cutting planes improve the solution time by an average of seven times over solving the original M-SMIP model without cutting planes. Our comparative analysis shows that the M-SMIP model outperforms six different heuristic approaches proposed for the management of EAB. Results from optimally solving our M-SMIP model imply that under a belief of infestation, it is critical to apply surveillance immediately to locate EAB and then prioritize treatment of minimally infested trees followed by removal of highly infested trees. Summary of Contributions: Emerald ash borer (EAB) is one of the most damaging invasive species ever to reach the United States, damaging millions of ash trees. Much of the economic impact of EAB occurs in cities, where high-value ash trees grow in abundance along streets and in yards and parks. This paper addresses the joint optimization of surveillance and control of the emerald ash borer invasion, which is a novel application for the INFORMS society because, to our knowledge, this specific problem of EAB management has not been published before in any OR/MS journals. We develop a new multi-stage stochastic mixed-integer programming (MSS-MIP) formulation, and we apply our model to surveillance and control of EAB in cities. Our MSS-MIP model aims to help city managers maximize the net benefits of their healthy ash trees by determining the optimal timing and target population for surveying, treating, and removing infested ash trees while taking into account the spatiotemporal stochastic growth of the EAB infestation. We develop a new cutting plane methodology motivated by our problem, which could also be applied to other stochastic MIPs. Our cutting plane approach provides significant computational benefit in solving the problem. Specifically, proposed cutting planes improve the solution time by an average of seven times over solving the original M-SMIP model without cutting planes. We calibrate and validate our model using seven years of ash infestation observations in forests near Toledo, Ohio. We then apply our model to an urban forest in Burnsville, Minnesota, that is threatened by EAB. Our results provide insights into the optimal timing and location of EAB surveillance and control strategies.
  • Learning Optimal Solutions via an LSTM-Optimization Framework
    Yilmaz, Dogacan; Büyüktahtakın, İ. Esra (Springer, 2023-06-06)
    In this study, we present a deep learning-optimization framework to tackle dynamic mixed-integer programs. Specifically, we develop a bidirectional Long Short Term Memory (LSTM) framework that can process information forward and backward in time to learn optimal solutions to sequential decision-making problems. We demonstrate our approach in predicting the optimal decisions for the single-item capacitated lot-sizing problem (CLSP), where a binary variable denotes whether to produce in a period or not. Due to the dynamic nature of the problem, the CLSP can be treated as a sequence labeling task where a recurrent neural network can capture the problem's temporal dynamics. Computational results show that our LSTM-Optimization (LSTM-Opt) framework significantly reduces the solution time of benchmark CLSP problems without much loss in feasibility and optimality. For example, the predictions at the 85% level reduce the CPLEX solution time by a factor of 9 on average for over 240,000 test instances with an optimality gap of less than 0.05% and 0.4% infeasibility in the test set. Also, models trained using shorter planning horizons can successfully predict the optimal solution of the instances with longer planning horizons. For the hardest data set, the LSTM predictions at the 25% level reduce the solution time of 70 CPU hours to less than 2 CPU minutes with an optimality gap of 0.8% and without any infeasibility. The LSTM-Opt framework outperforms classical ML algorithms, such as the logistic regression and random forest, in terms of the solution quality, and exact approaches, such as the (`, S) and dynamic programming-based inequalities, with respect to the solution time improvement. Our machine learning approach could be beneficial in tackling sequential decision-making problems similar to CLSP, which need to be solved repetitively, frequently, and in a fast manner.
  • Battery asset management with cycle life prognosis
    Liu, Xinyang; Zheng, Zhuoyuan; Büyüktahtakın, İ. Esra; Zhou, Zhi; Wang, Pingfeng (Elsevier, 2021-12)
    Battery Asset Management problem determines the minimum cost replacement schedules for each individual asset in a group of battery assets that operate in parallel. Battery cycle life varies under different operating conditions including temperature, depth of discharge, charge rate, etc., and a battery deteriorates due to usage, which cannot be handled by current asset management models. This paper presents a new battery asset management methodology where battery cycle life prognosis is integrated with parallel asset management to reduce lifecycle cost of the Battery Energy Storage Systems (BESS). For the battery failure time prognosis, a nonlinear physics-based battery capacity fade model is developed and incorporated in parallel asset management model to update battery capacity over time. Experiment results have shown that the developed battery asset management methodology can be conveniently used to facilitate BESS asset management decision making thereby decreasing asset lifecycle costs.
  • Optimizing surveillance and management of emerald ash borer in urban environments
    Bushaj, Sabah; Büyüktahtakın, İ. Esra; Yemshanov, Denys; Haight, Robert G. (Wiley, 2020-05-21)
    Emerald ash borer (EAB), a wood-boring insect native to Asia, was discovered near Detroit in 2002 and has spread and killed millions of ash trees throughout the eastern United States and Canada. EAB causes severe damage in urban areas where it kills high-value ash trees that shade streets, homes, and parks and costs homeowners and local governments millions of dollars for treatment, removal, and replacement of infested trees. We present a multistage, stochastic, mixed-integer programming model to help decision-makers maximize the public benefits of preserving healthy ash trees in an urban environment. The model allocates resources to surveillance of the ash population and subsequent treatment and removal of infested trees over time. We explore the multistage dynamics of an EAB outbreak with a dispersal mechanism and apply the optimization model to explore surveillance, treatment, and removal options to manage an EAB outbreak in Winnipeg, a city of Manitoba, Canada. Recommendation to Resource Managers Our approach demonstrates that timely detection and early response are critical factors for maximizing the number of healthy trees in urban areas affected by the pest outbreak. Treatment of the infested trees is most effective when done at the earliest stage of infestation. Treating asymptomatic trees at the earliest stages of infestation provides higher net benefits than tree removal or no-treatment options. Our analysis suggests the use of branch sampling as a more accurate method than the use of sticky traps to detect the infested asymptomatic trees, which enables treating and removing more infested trees at the early stages of infestation. Our results also emphasize the importance of allocating a sufficient budget for tree removal to manage emerald ash borer infestations in urban environments. Tree removal becomes a less useful option in small-budget solutions where the optimal policy is to spend most of the budget on treatments.
  • A non-anticipative learning-optimization framework for solving multi-stage stochastic programs
    Yilmaz, Dogacan; Büyüktahtakın, İ. Esra (Springer, 2024-07-03)
    We present a non-anticipative learning- and scenario-based prediction-optimization (ScenPredOpt) framework that combines deep learning, heuristics, and mathematical solvers for solving combinatorial problems under uncertainty. Specifically, we transform neural machine translation frameworks to predict the optimal solutions of scenario-based multi-stage stochastic programs. The learning models are trained efficiently using the input and solution data of the multi-stage single-scenario deterministic problems. Then our ScenPredOpt framework creates a mapping from the inputs used in training into an output of predictions that are close to optimal solutions. We present a Non-anticipative Encoder-Decoder with Attention (NEDA) approach, which ensures the non-anticipativity property of multi-stage stochastic programs and, thus, time consistency by calibrating the learned information based on the problem’s scenario tree and adjusting the hidden states of the neural network. In our ScenPredOpt framework, the percent predicted variables used for the solution are iteratively reduced through a relaxation of the problem to eliminate infeasibility. Then, a linear relaxation-based heuristic is performed to further reduce the solution time. Finally, a mathematical solver is used to generate the complete solution. We present the results on two NP-Hard sequential optimization problems under uncertainty: stochastic multi-item capacitated lot-sizing and stochastic multistage multidimensional knapsack. The results show that the solution time can be reduced by a factor of 599 with an optimality gap of only 0.08%. We compare the results of the ScenPredOpt framework with cutting-edge exact and heuristic solution algorithms for the problems studied and find that our framework is more effective. Additionally, the computational results demonstrate that ScenPredOpt can solve instances with a larger number of items and scenarios than the trained ones. Our non-anticipative learning-optimization approach can be beneficial for stochastic programming problems involving binary variables that are solved repeatedly with various types of dimensions and similar decisions at each period.
  • A K-means Supported Reinforcement Learning Framework to Multi-dimensional Knapsack
    Bushaj, Sabah; Büyüktahtakın, İ. Esra (Springer, 2024-02-15)
    In this paper, we address the difficulty of solving large-scale multi-dimensional knapsack instances (MKP), presenting a novel deep reinforcement learning (DRL) framework. In this DRL framework, we train different agents compatible with a discrete action space for sequential decision-making while still satisfying any resource constraint of the MKP. This novel framework incorporates the decision variable values in the 2D DRL where the agent is responsible for assigning a value of 1 or 0 to each of the variables. To the best of our knowledge, this is the first DRL model of its kind in which a 2D environment is formulated, and an element of the DRL solution matrix represents an item of the MKP. Our framework is configured to solve MKP instances of different dimensions and distributions. We propose a K-means approach to obtain an initial feasible solution that is used to train the DRL agent. We train four different agents in our framework and present the results comparing each of them with the CPLEX commercial solver. The results show that our agents can learn and generalize over instances with different sizes and distributions. Our DRL framework shows that it can solve medium-sized instances at least 45 times faster in CPU solution time and at least 10 times faster for large instances, with a maximum solution gap of 0.28% compared to the performance of CPLEX. Furthermore, at least 95% of the items are predicted in line with the CPLEX solution. Computations with DRL also provide a better optimality gap with respect to state-of-the-art approaches.
  • High-Quality Dataset-Sharing and Trade Based on A Performance-Oriented Directed Graph Neural Network
    Zeng, Yingyan; Zhou, Xiaona; Chilukuri, Premith; Lourentzou, Ismini; Jin, Ran (2025)
    The advancement of Artificial Intelligence (AI) models heavily relies on large high-quality datasets. However, in advanced manufacturing, collecting such data is time-consuming and labor-intensive for a single enterprise. Hence, it is important to establish a context-aware and privacy-preserving data sharing system to share small-but-high-quality datasets between trusted stakeholders. Existing data sharing approaches have explored privacy-preserving data distillation methods and focused on valuating individual samples tied to a specific AI model, limiting their flexibility across data modalities, AI tasks, and dataset ownership. In this work, we propose a performance-oriented representation learning (PORL) framework in a Directed Graph Neural Network (DiGNN). PORL distills raw datasets into privacy-preserving proxy datasets for sharing and learns compact meta data representations for each stakeholder locally. The meta data will then be used in DiGNN to forecast the AI model performance and guide the sharing via graph-level supervised learning. The effectiveness of the PORL-DiGNN is validated by two case studies: data sharing in the semiconducting manufacturing network between similar processes to create similar quality defect models; and data sharing in the design and manufacturing network of Microbial Fuel Cell anodes between upstream (design) and downstream (Additive Manufacturing) stages to create distinct but related AI models.
  • Data-Driven Sample Average Approximation with Covariate Information
    Kannan, Rohit; Bayraksan, Guezin; Luedtke, James R. (INFORMS, 2025-01-06)
    We study optimization for data-driven decision-making when we have observations of the uncertain parameters within an optimization model together with concurrent observations of covariates. The goal is to choose a decision that minimizes the expected cost conditioned on a new covariate observation. We investigate two data-driven frameworks that integrate a machine learning prediction model within a stochastic programming sample average approximation (SAA) for approximating the solution to this problem. One SAA framework is new and uses leave-one-out residuals for scenario generation. The frameworks we investigate are flexible and accommodate parametric, nonparametric, and semiparametric regression techniques. We derive conditions on the data generation process, the prediction model, and the stochastic program under which solutions of these data-driven SAAs are consistent and asymptotically optimal, and also derive finite sample guarantees. Computational experiments validate our theoretical results, demonstrate examples where our datadriven formulations have advantages over existing approaches (even if the prediction model is misspecified), and illustrate the benefits of our data-driven formulations in the limited data regime.
  • Residuals-based distributionally robust optimization with covariate information
    Kannan, Rohit; Bayraksan, Guezin; Luedtke, James R. (Springer, 2023-09-26)
    We consider data-driven approaches that integrate a machine learning prediction model within distributionally robust optimization (DRO) given limited joint observations of uncertain parameters and covariates. Our framework is flexible in the sense that it can accommodate a variety of regression setups and DRO ambiguity sets. We investigate asymptotic and finite sample properties of solutions obtained using Wasserstein, sample robust optimization, and phi-divergence-based ambiguity sets within our DRO formulations, and explore cross-validation approaches for sizing these ambiguity sets. Through numerical experiments, we validate our theoretical results, study the effectiveness of our approaches for sizing ambiguity sets, and illustrate the benefits of our DRO formulations in the limited data regime even when the prediction model is misspecified.