Search

Diagnostic errors are common and can result in serious patient harm. Making the right diagnosis often requires significant diagnostic effort. Yet most physician payment schemes are procedure based and do not account for diagnostic effort or accuracy because of observability issues. In this paper, we develop a parsimonious model to examine the impact of a physician payment scheme on a physician's decisions to (1) exert diagnostic effort and (2) perform a confirmatory test. High effort provides an informative (though imperfect) signal of the patient's true state; the test is confirmatory in that it is a prerequisite for diagnosing a severe condition. Our model uses a two-step diagnostic process to capture the interaction between the physician's diagnostic effort and testing decisions. We show that under a fee-for-service payment scheme, the physician may view the diagnostic effort and the confirmatory test as either complementary or substitutive, depending on the additional revenue from testing. We also reveal nonmonotonic properties such that a more patient-centered physician may not exert more effort or provide a more accurate diagnosis. In addition, either a flat or differentiated payment scheme may be optimal. We also show that an alternative payment scheme, under which the revenue from the confirmatory test is contingent on its result, can induce the social optimum under certain conditions. With the advent of artificial intelligence as part of the standard of care and its increasing use as confirmatory tests, our research has implications for the design of physician payment systems in light of concerns about the potential erosion of individual attention. This paper was accepted by Jayashankar Swaminathan, operations management. Funding: T. Dai was supported by the Johns Hopkins Discovery Award (2022–2024) and the Hopkins Business of Health Initiative Seed Grant (2022–2023). Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2023.4937. [ABSTRACT FROM AUTHOR]

The rapid adoption of artificial intelligence (AI) technologies by many organizations has recently raised concerns that AI may eventually replace humans in certain tasks. In fact, when used in collaboration, machines can significantly enhance the complementary strengths of humans. Indeed, because of their immense computing power, machines can perform specific tasks with incredible accuracy. In contrast, human decision makers (DMs) are flexible and adaptive but constrained by their limited cognitive capacity. This paper investigates how machine-based predictions may affect the decision process and outcomes of a human DM. We study the impact of these predictions on decision accuracy, the propensity and nature of decision errors, and the DM's cognitive efforts. To account for both flexibility and limited cognitive capacity, we model the human decision-making process in a rational inattention framework. In this setup, the machine provides the DM with accurate but sometimes incomplete information at no cognitive cost. We fully characterize the impact of machine input on the human decision process in this framework. We show that machine input always improves the overall accuracy of human decisions but may nonetheless increase the propensity of certain types of errors (such as false positives). The machine can also induce the human to exert more cognitive efforts, although its input is highly accurate. Interestingly, this happens when the DM is most cognitively constrained, for instance, because of time pressure or multitasking. Synthesizing these results, we pinpoint the decision environments in which human-machine collaboration is likely to be most beneficial. This paper was accepted by Jeannette Song, operations management. Supplemental Material: The data files and online appendices are available at https://doi.org/10.1287/mnsc.2023.4744. [ABSTRACT FROM AUTHOR]

This paper studies an emerging subscription model called ship-then-shop. Leveraging its predictive analytics and artificial intelligence (AI) capability, the ship-then-shop firm curates and ships a product to the consumer, after which the consumer shops (i.e., evaluates product fit and makes a purchase decision). The consumer first pays the up-front ship-then-shop subscription fee prior to observing product fit and then pays the product price afterward if the consumer decides to purchase. We investigate how the firm balances the subscription fee and product price to maximize its profit when consumers can showroom. A key finding is the ship-then-shop firm's nonmonotonic surplus extraction strategy with respect to its prediction capability. As prediction capability increases, the firm first switches from ex ante to ex post surplus extraction (by lowering fees and raising prices). However, if the prediction capability increases further, the firm reverts to ex ante surplus extraction (by raising fees and capping prices). We also find that the ship-then-shop model is most profitable when (i) the prediction capability is advanced, (ii) the search friction in the market is large, or (iii) the product match potential is large. Finally, we show that the marginal return of AI capability on the firm's profit decreases in search friction but increases in product match potential. Taken together, we provide managerially relevant insights to help guide the implementation of the innovative subscription model. This paper was accepted by Dmitri Kuksov, marketing. Supplemental Material: The online appendix is available at https://doi.org/10.1287/mnsc.2023.4723. [ABSTRACT FROM AUTHOR]

Through repeated interactions, firms today refine their understanding of individual users' preferences adaptively for personalization. In this paper, we use a continuous-time bandit model to analyze firms that recommend content to multihoming consumers, a representative setting for strategic learning of consumer preferences to maximize lifetime value. In both monopoly and duopoly settings, we compare a forward-looking recommendation algorithm that balances exploration and exploitation to a myopic algorithm that only maximizes the quality of the next recommendation. Our analysis shows that, compared with a monopoly, firms competing for users' attention focus more on exploitation than exploration. When users are impatient, competition decreases the return from developing a forward-looking algorithm. In contrast, development of a forward-looking algorithm may hurt users under monopoly but always benefits users under competition. Competing firms' decisions to invest in a forward-looking algorithm can create a prisoner's dilemma. Our results have implications for artificial intelligence adoption and for policy makers on the effect of market power on innovation and consumer welfare. This paper was accepted by Dmitri Kuksov, marketing. Supplemental Material: The online appendix is available at https://doi.org/10.1287/mnsc.2023.4722. [ABSTRACT FROM AUTHOR]

As artificial intelligence (AI) applications become more pervasive, it is critical to understand how knowledge workers with different levels and types of experience can team with AI for productivity gains. We focus on the influence of two major types of human work experience (narrow experience based on the specific task volume and broad experience based on seniority) on the human-AI team dynamics. We developed an AI solution for medical chart coding in a publicly traded company and conducted a field study among the knowledge workers. Based on a detailed analysis performed at the medical chart level, we find evidence that AI benefits workers with greater task-based experience, but senior workers gain less from AI than their junior colleagues. Further investigation reveals that the relatively lower productivity lift from AI is not a result of seniority per se but lower trust in AI, likely triggered by the senior workers' broader job responsibilities. This study provides new empirical insights into the differential roles of worker experience in the collaborative dynamics between AI and knowledge workers, which have important societal and business implications. This paper was accepted by Kartik Hosanagar, information systems. Funding: This work was supported by Inovalon [Sponsor of the Health Insights AI Laboratory]. Supplemental Material: The data files and online appendix are available at https://doi.org/10.1287/mnsc.2021.00588. [ABSTRACT FROM AUTHOR]

Decision making on the job is becoming increasingly important in the labor market, in which there is an unprecedented rise in demand for workers with problem-solving and critical-thinking skills. This paper investigates how indoor air quality affects the quality of strategic decision making based on data from official chess tournaments. Our main analysis relies on a unique data set linking the readings of air-quality monitors inside the tournament room to the quality of 30,000 moves, each of them objectively evaluated by a powerful artificial intelligence–based chess engine. The results show that poor indoor air quality hampers players' decision making. We find that an increase in the indoor concentration of fine particulate matter (PM 2.5 ) by 10 μ g / m 3 (corresponding to 75% of a standard deviation in our sample) increases a player's probability of making an erroneous move by 26.3%. The decomposition of the effects by different stages of the game shows that time pressure amplifies the damage of poor air quality to the players' decisions. We implement a number of robustness checks and conduct a replication exercise with analogous move-quality data from games in the top national league showing the strength of our results. The results highlight the costs of poor air quality for highly skilled professionals faced with strategic decisions under time pressure. This paper was accepted by Prof. Yan Chen, behavioral economics and decision analysis. Funding: The authors gratefully acknowledge the financial support from the Graduate School of Business and Economics at Maastricht University as well as the Institute of Labor Economics Bonn. Supplemental Material: The online appendix and data are available at https://doi.org/10.1287/mnsc.2022.4643. [ABSTRACT FROM AUTHOR]

Because the average treatment effect (ATE) measures the change in social welfare, even if positive, there is a risk of negative effect on, say, some 10% of the population. Assessing such risk is difficult, however, because any one individual treatment effect (ITE) is never observed, so the 10% worst-affected cannot be identified, whereas distributional treatment effects only compare the first deciles within each treatment group, which does not correspond to any 10% subpopulation. In this paper, we consider how to nonetheless assess this important risk measure, formalized as the conditional value at risk (CVaR) of the ITE distribution. We leverage the availability of pretreatment covariates and characterize the tightest possible upper and lower bounds on ITE-CVaR given by the covariate-conditional average treatment effect (CATE) function. We then proceed to study how to estimate these bounds efficiently from data and construct confidence intervals. This is challenging even in randomized experiments as it requires understanding the distribution of the unknown CATE function, which can be very complex if we use rich covariates to best control for heterogeneity. We develop a debiasing method that overcomes this and prove it enjoys favorable statistical properties even when CATE and other nuisances are estimated by black box machine learning or even inconsistently. Studying a hypothetical change to French job search counseling services, our bounds and inference demonstrate a small social benefit entails a negative impact on a substantial subpopulation. This paper was accepted by J. George Shanthikumar, data science. Funding: This work was supported by the Division of Information and Intelligent Systems [Grant 1939704]. Supplemental Material: The data files and online appendices are available at https://doi.org/10.1287/mnsc.2023.4819. [ABSTRACT FROM AUTHOR]

Drawing on the notion of compensatory behavior, this paper studies how students compensate for learning loss during a pandemic and what role artificial intelligence (AI) plays in this regard. We further probe into a difference in compensatory behavior for learning loss in terms of quantity, pattern, and pace (i.e., tripartite aspect of learning behavior) of AI-powered learning app usage depending on the level of pandemic threat and the proximity of a goal to students. Results show that the pandemic threat affects student learning behavior differently. Immediately following the COVID-19 outbreak, students who live in the epicenter of the outbreak (versus those who do not) use the app less at first, but with time, they use it more (quantity), on a more regular basis (pattern), and rebound to a curriculum path (pace) comparable to students who do not live in the outbreak's epicenter. These findings collectively explain behavior that is consistent with compensation for learning loss. The results also partially corroborate the goal-proximity effect, revealing that proximity to a goal (e.g., the degree to which the national university admission exam is approaching) has a moderating role in explaining the tripartite perspective of student learning behavior. Overall, these findings have important theoretical and practical implications for understanding how innovative education technologies can not only facilitate student learning during adversity, but also support learning recovery after adversity. This paper was accepted by D. J. Wu, information systems. Supplemental Material: Data files available at https://doi.org/10.1287/mnsc.2022.4531. [ABSTRACT FROM AUTHOR]

We initiate the study of incentive-compatible forecasting competitions in which multiple forecasters make predictions about one or more events and compete for a single prize. We have two objectives: (1) to incentivize forecasters to report truthfully and (2) to award the prize to the most accurate forecaster. Proper scoring rules incentivize truthful reporting if all forecasters are paid according to their scores. However, incentives become distorted if only the best-scoring forecaster wins a prize, since forecasters can often increase their probability of having the highest score by reporting more extreme beliefs. In this paper, we introduce two novel forecasting competition mechanisms. Our first mechanism is incentive compatible and guaranteed to select the most accurate forecaster with probability higher than any other forecaster. Moreover, we show that in the standard single-event, two-forecaster setting and under mild technical conditions, no other incentive-compatible mechanism selects the most accurate forecaster with higher probability. Our second mechanism is incentive compatible when forecasters' beliefs are such that information about one event does not lead to belief updates on other events, and it selects the best forecaster with probability approaching one as the number of events grows. Our notion of incentive compatibility is more general than previous definitions of dominant strategy incentive compatibility in that it allows for reports to be correlated with the event outcomes. Moreover, our mechanisms are easy to implement and can be generalized to the related problems of outputting a ranking over forecasters and hiring a forecaster with high accuracy on future events. This paper was accepted by Yan Chen, behavioral economics and decision analysis. Funding: This work was supported by the European Research Council [Grant ERC StG 307036] and the National Science Foundation [Grant CCF-1445755]. [ABSTRACT FROM AUTHOR]

This paper explores how firms that lack expertise in machine learning (ML) can leverage the so-called AI Flywheel effect. This effect designates a virtuous cycle by which as an ML product is adopted and new user data are fed back to the algorithm, the product improves, enabling further adoptions. However, managing this feedback loop is difficult, especially when the algorithm is contracted out. Indeed, the additional data that the AI Flywheel effect generates may change the provider's incentives to improve the algorithm over time. We formalize this problem in a simple two-period moral hazard framework that captures the main dynamics among ML, data acquisition, pricing, and contracting. We find that the firm's decisions crucially depend on how the amount of data on which the machine is trained interacts with the provider's effort. If this effort has a more (less) significant impact on accuracy for larger volumes of data, the firm underprices (overprices) the product. Interestingly, these distortions sometimes improve social welfare, which accounts for the customer surplus and profits of both the firm and provider. Further, the interaction between incentive issues and the positive externalities of the AI Flywheel effect has important implications for the firm's data collection strategy. In particular, the firm can boost its profit by increasing the product's capacity to acquire usage data only up to a certain level. If the product collects too much data per user, the firm's profit may actually decrease (i.e., more data are not necessarily better). This paper was accepted by Jayashankar Swaminathan, operations management. Supplemental Material: The data files and e-companion are available at https://doi.org/10.1287/mnsc.2022.4333. [ABSTRACT FROM AUTHOR]

The trust game, a simple two-player economic exchange, is extensively used as an experimental measure for trust and trustworthiness of individuals. We construct deep neural network–based artificial intelligence (AI) agents to participate a series of experiments based upon the trust game. These artificial agents are trained by playing with one another repeatedly without any prior knowledge, assumption, or data regarding human behaviors. We find that, under certain conditions, AI agents produce actions that are qualitatively similar to decisions of human subjects reported in the trust game literature. Factors that influence the emergence and levels of cooperation by artificial agents in the game are further explored. This study offers evidence that AI agents can develop trusting and cooperative behaviors purely from an interactive trial-and-error learning process. It constitutes a first step to build multiagent-based decision support systems in which interacting artificial agents are capable of leveraging social intelligence to achieve better outcomes collectively. This paper was accepted by Yan Chen, behavioral economics and decision analysis. Funding: Y. (D.) Wu extends her gratitude for the financial support provided through the RSCA Seed [Grant 22-RSG-01-004] from the San Jose State University. Supplemental Material: Data are available at https://doi.org/10.1287/mnsc.2023.4782. [ABSTRACT FROM AUTHOR]

Ensuring fairness in algorithmic decision making is a crucial policy issue. Current legislation ensures fairness by barring algorithm designers from using demographic information in their decision making. As a result, to be legally compliant, the algorithms need to ensure equal treatment. However, in many cases, ensuring equal treatment leads to disparate impact particularly when there are differences among groups based on demographic classes. In response, several "fair" machine learning (ML) algorithms that require impact parity (e.g., equal opportunity) at the cost of equal treatment have recently been proposed to adjust for the societal inequalities. Advocates of fair ML propose changing the law to allow the use of protected class-specific decision rules. We show that the proposed fair ML algorithms that require impact parity, while conceptually appealing, can make everyone worse off, including the very class they aim to protect. Compared with the current law, which requires treatment parity, the fair ML algorithms, which require impact parity, limit the benefits of a more accurate algorithm for a firm. As a result, profit maximizing firms could underinvest in learning, that is, improving the accuracy of their machine learning algorithms. We show that the investment in learning decreases when misclassification is costly, which is exactly the case when greater accuracy is otherwise desired. Our paper highlights the importance of considering strategic behavior of stake holders when developing and evaluating fair ML algorithms. Overall, our results indicate that fair ML algorithms that require impact parity, if turned into law, may not be able to deliver some of the anticipated benefits. This paper was accepted by Kartik Hosanagar, information systems. [ABSTRACT FROM AUTHOR]

Managing perishable inventory systems with positive lead times and finite ordering capacities is important but notoriously difficult in both theory and computation. The optimal control policy is extremely complicated, and no effective heuristic policy has been proposed in the literature. In this paper, we develop an easy-to-compute approximation algorithm for this class of problems and prove that it admits a theoretical worst-case performance guarantee under independent and many commonly used positively correlated demand processes. Our worst-case analysis departs significantly from those in the previous studies, requiring several novel ideas. In particular, we introduce a transient unit-matching rule to dynamically match the supply and demand units, and the notion of associated demand processes that provides the right future demand information to establish the desired results. Our numerical study demonstrates the effectiveness of the proposed algorithm. This paper was accepted by Yinyu Ye, optimization. [ABSTRACT FROM AUTHOR]

Motivated by operations research applications, such as inventory control and real-time bidding, we consider undiscounted reinforcement learning in Markov decision processes under model uncertainty and temporal drifts. In this setting, both the latent reward and state transition distributions are allowed to evolve over time, as long as their respective total variations, quantified by suitable metrics, do not exceed certain variation budgets. We first develop the sliding window upper confidence bound for reinforcement learning with confidence-widening (SWUCRL2-CW) algorithm and establish its dynamic regret bound when the variation budgets are known. In addition, we propose the bandit-over-reinforcement learning algorithm to adaptively tune the SWUCRL2-CW algorithm to achieve the same dynamic regret bound but in a parameter-free manner (i.e., without knowing the variation budgets). Finally, we conduct numerical experiments to show that our proposed algorithms achieve superior empirical performance compared with existing algorithms. Notably, under nonstationarity, historical data samples may falsely indicate that state transition rarely happens. This thus presents a significant challenge when one tries to apply the conventional optimism in the face of uncertainty principle to achieve a low dynamic regret bound. We overcome this challenge by proposing a novel confidence-widening technique that incorporates additional optimism into our learning algorithms. To extend our theoretical findings, we demonstrate, in the context of single-item inventory control with lost sales, fixed cost, and zero lead time, how one can leverage special structures on the state transition distributions to achieve improved dynamic regret bound in time-varying demand environments. This paper was accepted by J. George Shanthikumar, data science. Funding: The authors acknowledge support from the Massachusetts Institute of Technology (MIT) Data Science Laboratory and the MIT–IBM partnership in artificial intelligence. W. C. Cheung acknowledges support from the Singapore Ministry of Education [Tier 2 Grant MOE-T2EP20121-0012]. Supplemental Material: The data files and online appendix are available at https://doi.org/10.1287/mnsc.2023.4704. [ABSTRACT FROM AUTHOR]

Strategic decisions are increasingly delegated to algorithms. We extend previous results of the algorithmic collusion literature to the context of dynamic optimization with imperfect monitoring by analyzing a setting where a limited number of agents use simple and independent machine-learning algorithms to buy and sell a storable good. No specific instruction is given to them, only that their objective is to maximize profits based solely on past market prices and payoffs. With an original application to battery operations, we observe that the algorithms learn quickly to reach seemingly collusive decisions, despite the absence of any formal communication between them. Building on the findings of the existing literature on algorithmic collusion, we show that seeming collusion could originate in imperfect exploration rather than excessive algorithmic sophistication. We then show that a regulator may succeed in disciplining the market to produce socially desirable outcomes by enforcing decentralized learning or with adequate intervention during the learning process. This paper was accepted by Gabriel Weintraub, revenue management and market analytics. Supplemental Material: The data files are available at https://doi.org/10.1287/mnsc.2022.4623. [ABSTRACT FROM AUTHOR]

Deep reinforcement learning (DRL) has opened up many unprecedented opportunities in revolutionizing the digital marketing field. In this study, we designed a DRL-based personalized targeting strategy in a sequential setting. We show that the strategy is able to address three important challenges of sequential targeting: (1) forward looking (balancing between a firm's current revenue and future revenues), (2) earning while learning (maximizing profits while continuously learning through exploration-exploitation), and (3) scalability (coping with a high-dimensional state and policy space). We illustrate this through a novel design of a DRL-based artificial intelligence (AI) agent. To better adapt DRL to complex consumer behavior dimensions, we proposed a quantization-based uncertainty learning heuristic for efficient exploration-exploitation. Our policy evaluation results through simulation suggest that the proposed DRL agent generates 26.75% more long-term revenues than can the non-DRL approaches on average and learns 76.92% faster than the second fastest model among all benchmarks. Further, in order to better understand the potential underlying mechanisms, we conducted multiple interpretability analyses to explain the patterns of learned optimal policy at both the individual and population levels. Our findings provide important managerial-relevant and theory-consistent insights. For instance, consecutive price promotions at the beginning can capture price-sensitive consumers' immediate attention, whereas carefully spaced nonpromotional "cooldown" periods between price promotions can allow consumers to adjust their reference points. Additionally, consideration of future revenues is necessary from a long-term horizon, but weighing the future too much can also dampen revenues. In addition, analyses of heterogeneous treatment effects suggest that the optimal promotion sequence pattern highly varies across the consumer engagement stages. Overall, our study results demonstrate DRL's potential to optimize these strategies' combination to maximize long-term revenues. This paper was accepted by Kartik Hosanagar, information systems. Supplemental Material: The data files and online appendix are available at https://doi.org/10.1287/mnsc.2022.4621. [ABSTRACT FROM AUTHOR]

In a contest in which solvers with heterogeneous expertise exert effort to develop solutions, a recent paper [Terwiesch C, Xu Y (2008) Innovation contests, open innovation, and multiagent problem solving. Management Sci. 54(9):1529-1543] argues that as more solvers enter the contest, every solver will reduce effort due to a lower probability of winning the contest. This paper corrects mistakes in this theory, and shows that there exist high-expertise solvers who may raise their effort in response to increased competition. This is because more entrants raise the expected best performance among other solvers, creating positive incentives for solvers to exert higher effort to win the contest. Because of this positive effect, we find that a free-entry open contest is more likely to be optimal to a contest organizer than what Terwiesch and Xu (2008) and other prior literature asserted. [ABSTRACT FROM AUTHOR]

Switchback experiments, where a firm sequentially exposes an experimental unit to random treatments, are among the most prevalent designs used in the technology sector, with applications ranging from ride-hailing platforms to online marketplaces. Although practitioners have widely adopted this technique, the derivation of the optimal design has been elusive, hindering practitioners from drawing valid causal conclusions with enough statistical power. We address this limitation by deriving the optimal design of switchback experiments under a range of different assumptions on the order of the carryover effect—the length of time a treatment persists in impacting the outcome. We cast the optimal experimental design problem as a minimax discrete optimization problem, identify the worst-case adversarial strategy, establish structural results, and solve the reduced problem via a continuous relaxation. For switchback experiments conducted under the optimal design, we provide two approaches for performing inference. The first provides exact randomization-based p-values, and the second uses a new finite population central limit theorem to conduct conservative hypothesis tests and build confidence intervals. We further provide theoretical results when the order of the carryover effect is misspecified and provide a data-driven procedure to identify the order of the carryover effect. We conduct extensive simulations to study the numerical performance and empirical properties of our results and conclude with practical suggestions. This paper was accepted by George Shanthikumar, big data analytics. Funding: The authors thank the Massachusetts Institute of Technology (MIT)-IBM partnership in Artificial Intelligence and the MIT Data Science Laboratory for support. Supplemental Material: Data and the online appendix are available at https://doi.org/10.1287/mnsc.2022.4583. [ABSTRACT FROM AUTHOR]

Expert systems often employ a weight on rules to capture conditional probabilities. For example, in classic rule-based settings, Pr(h|e) = x is used to mean "If e is known to be true then conclude h is true with probability x." Further, other probability-based approaches, such as influence diagrams and Bayes' Nets are increasingly being used to support decision making through decision support systems. Although algorithms for these systems have received substantial attention, less attention has been given to knowledge acquisition of probabilities used in these systems. However, the underlying probabilities are critical because they lead the user to particular solutions. Accordingly, the purpose of this paper is to investigate the quality of probability knowledge when it is acquired from groups or individuals.This paper summarizes the results of an empirical cognitive study on the ability of individuals and groups to provide consistent sets of probabilities Pr(A), Pr(B), Pr(A|B) and Pr(B|A). The analysis of these probabilities allowed the study of the ability of subjects to account for Bayes' theorem reversals, a basic assumption made by virtually all algorithms. It was found that knowledge acquisition from groups provided more correct orderings to the probabilities than knowledge acquisition from individuals. This suggests that knowledge acquisition from groups is more likely to obtain correct probability knowledge. [ABSTRACT FROM AUTHOR]

AI automated plan synthesis programs ("planners") typically represent plans as a partially ordered network whose nodes are instants in time and whose arcs are precedence constraints. Such representations arc essentially PERT charts. This paper provides an introduction to AI planners and illustrates the tact that many such PERT chart representations are created by a planner as it attempts to find a suitable plan. A planner's search must he guided if it is to produce suitable plans in realistic problem domains. One form of guidance is the imposition of constraints which must be satisfied by a valid plan. This paper concentrates on representing and reasoning about such constraints. Successful planning in some domains, e.g. job-shop scheduling, requires the recognition and satisfaction of constraints which prevent the simultaneous use of a shared resource by multiple agents. Such requirements can he expressed as disjunctive precedence constraints. In addition, the search for a valid plan is often structured so that rough plans are successively refined into plans with increased detail. In this refinement process, new constraints are often discovered. It should therefore be easy to add such new constraints. Finally, it is helpful to be able to express upper bound constraints on the elapsed time between two nodes in a PERT chart representation of a plan. Thus, this paper reports on knowledge representation and reasoning processes which facilitate: (1) disjunctive constraints, (2) the introduction of new constraints, and (3) upper bound constraints. A planner must backtrack when it is evident that further refinement of its current skeletal plan would be fruitless. Such an inconsistent state can be recognized when a plan's PERT chart representation contains an illegal cycle of positive overall length. When backtracking is required, dependency-directed backtracking is facilitated by recording "Nogood sets" of arcs whose simultaneous presence would result in such an illegal cycle. By using dependency-directed backtracking, chronological backtracking can be avoided. [ABSTRACT FROM AUTHOR]

We develop a data-driven decision model to improve process quality in manufacturing. A challenge for traditional methods in quality management is to handle high-dimensional and nonlinear manufacturing data. We address this challenge by adapting explainable artificial intelligence to the context of quality management. Specifically, we propose the use of nonlinear modeling with Shapley additive explanations to infer how a set of production parameters and the process quality of a manufacturing system are related. Thereby, we contribute a measure of process importance based on which manufacturers can prioritize processes for quality improvement. Grounded in quality management theory, our decision model selects improvement actions that target the sources of quality variation. The decision model is validated in a real-world application at a leading manufacturer of high-power semiconductors. Seeking to improve production yield, we apply our decision model to select improvement actions for a transistor chip product. We then conduct a field experiment to confirm the effectiveness of the improvement actions. Compared with the average yield in our sample, the experiment returns a reduction in yield loss of 21.7%. Furthermore, we report on results from a postexperimental rollout of the decision model, which also resulted in significant yield improvements. We demonstrate the operational value of explainable artificial intelligence by showing that critical drivers of process quality can go undiscovered by the use of traditional methods. This paper was accepted by Charles Corbett, operations management. [ABSTRACT FROM AUTHOR]

A "problem solver" (PS) is an agent who when interacting with other agents does not "put himself in their shoes" but rather chooses a best response to a uniform distribution over all possible configurations consistent with the information he receives about the other agents' moves. We demonstrate the special features of a PS by analyzing a modified coordination game. In the first stage, each of the other participants—who are treated as conventional players—chooses a location. The PS then receives some partial information about their moves and chooses his location. The PS wishes to coordinate with any one of the conventional players and they wish to coordinate with him but not with each other. Equilibria are characterized and shown to have different properties than those of Nash equilibria when the PS is treated as a conventional player. This paper was accepted by Joshua Gans, business strategy. [ABSTRACT FROM AUTHOR]

We build a game-theoretic model to examine how better demand forecasting resulting from algorithms, machine learning, and artificial intelligence affects the sustainability of collusion in an industry. We find that, although better forecasting allows colluding firms to better tailor prices to demand conditions, it also increases each firm's temptation to deviate to a lower price in time periods of high predicted demand. Overall, our research suggests that, despite concerns expressed by policy makers, better forecasting and algorithms can lead to lower prices and higher consumer surplus. This paper was accepted by Joshua Gans, business strategy. [ABSTRACT FROM AUTHOR]

Intelligent problem-solving, whether by man or by machine, implies selective rather than just rapid behavior. Humans achieve this selectivity through heuristics--principles that, on the average, contribute to reduction of search in problem-solving. Heuristic programming is the construction of computer problem-solving programs whose behavior is similarly organized. This paper surveys the present and potential use of heuristic programming in management science. [ABSTRACT FROM AUTHOR]

Information about the future may be instrumentally useful yet scary. For example, many patients shy away from precise genetic tests about their dispositions for severe diseases. They are afraid that a bad test result could render them desperate as a result of anticipatory feelings. We show that partially revealing tests are typically optimal when anticipatory utility interacts with an instrumental need for information. The same result emerges when patients rely on probability weighting. Optimal tests provide only two signals, which renders them easily implementable. While the good signal is typically precise, the bad one remains coarse. This way, patients have a substantial chance to learn that they are free of the genetic risk in question. Yet even if the test outcome is bad, they do not end in a situation without hope. This paper was accepted by Manel Baucells, decision analysis. [ABSTRACT FROM AUTHOR]

Web-based firms often rely on cloud-based computational resources to serve customers, but the number of customers they will serve is rarely known at the time of product launch. A recent innovation in cloud computing, known as autoscaling, allows companies to automatically scale their computational load up or down to match customer demand. We build a game theory model to examine how autoscaling will affect firms' decisions to enter a new market and the resulting equilibrium prices, profitability, and consumer surplus. The model produces novel results depending on the likelihood of a firm's success in the new market, differentiation among potential entrants, and the cost of computational capacity. For instance, in contrast to previous capacity commitment models with demand certainty, we show that autoscaling can mitigate price competition if the likelihood of a firm's success in the market is moderate and the cost of capacity is sufficiently low. This is because without autoscaling, the firms' uncertainty about demand would lead to excessive computational capacities and thus aggressive price competition. We also find that when the likelihood of success is sufficiently high, autoscaling facilitates entry for one firm yet deters a second firm from simultaneously entering the market. The online appendix is available at https://doi.org/10.1287/mnsc.2017.2891. This paper was accepted by Juanjuan Zhang, marketing. [ABSTRACT FROM AUTHOR]

Many probability forecasts are revised as new information becomes available, generating a time series of forecasts for a single event. Although methods for evaluating probability forecasts have been extensively studied, they apply to a single forecast per event. This paper is the first to evaluate probability forecasts that are made--and therefore revised--at many lead times for a single event. I postulate a norm for multi-period probability-forecasting systems and derive properties that should hold regardless of the forecasting process. I use these properties to develop methods for evaluating a forecasting system based on a sample. I apply these methods to the National Hurricane Center's wind-speed probability forecasts and to statistical election forecasts, finding evidence that both can be improved using the current set of predictors. [ABSTRACT FROM AUTHOR]

A skill's value depends on dynamic market conditions. To remain marketable, contractors need to keep reskilling themselves continuously. But choosing new skills to learn is an inherently hard task: Contractors have very little information about current and future market conditions, which often results in poor learning choices. Recommendation frameworks could reduce uncertainty in learning choices. However, conventional approaches would likely be inefficient; they would model previous (often poor) observed contractor learning behaviors to provide future career path recommendations while ignoring current market trends. This work proposes a framework that combines reinforcement learning, Bayesian inference, and gradient boosting to provide recommendations on how contractors should behave when choosing new skills to learn. Compared with standard recommender systems, this framework does not learn from previous (often poor) behaviors to make future recommendations. Instead, it relies on a Markov decision process to operate on a graph of feasible actions and dynamically recommend profitable career paths. The framework uses market information to identify current trends and project future wages. Based on this information, it recommends feasible, relevant actions that a contractor can take to learn new, in-demand skills. Evaluation of the framework on 1.73 million job applications from an online labor market shows that its implementation could increase (1) the marketplace's revenue by up to 6%, (2) contractors' wages by 22%, and (3) the diversity of new skill acquisitions by 47%. A comparison with alternative recommender systems highlights the limitations of approaches that make recommendations based on previously observed learning behaviors. This paper was accepted by Chris Forman, information systems. [ABSTRACT FROM AUTHOR]

Participatory budgeting enables the allocation of public funds by collecting and aggregating individual preferences. It has already had a sizable real-world impact, but making the most of this new paradigm requires rethinking some of the basics of computational social choice, including the very way in which individuals express their preferences. We attempt to maximize social welfare by using observed votes as proxies for voters' unknown underlying utilities, and analytically compare four preference elicitation methods: knapsack votes, rankings by value or value for money, and threshold approval votes. We find that threshold approval voting is qualitatively superior, and also performs well in experiments using data from real participatory budgeting elections. This paper was accepted by Yan Chen, decision analysis. [ABSTRACT FROM AUTHOR]

In this paper, we propose a new interactive procedure for solving multiple objective programming problems. Based upon feed-forward artificial neural networks (FFANNs), the method is called the Interactive FFANN Procedure. In the procedure, the decision maker articulates preference information over representative samples from the nondominated set either by assigning preference "values" to the sample solutions or by making pairwise comparisons in a fashion similar to that in the Analytic Hierarchy Process. With this information, a FFANN is trained to represent the decision maker's preference structure. Then, using the FFANN, an optimization problem is solved to search for improved solutions. An example is given to illustrate the Interactive FFANN Procedure. Also, the procedure is compared computationally with the Tchebycheff Method (Steuer and Choo 1983). The computational results indicate that the Interactive FFANN Procedure produces good solutions and is robust with regard to the neural network architecture. [ABSTRACT FROM AUTHOR]

With rapidly growing interest in the development of knowledge-based computer consulting systems for various problem domains, the difficulties associated with knowledge acquisition have special importance. This paper reports on the results of experiments designed to assess the effectiveness of an inductive algorithm in discovering predictive knowledge structures in financial data. The quality of the results are evaluated by comparing them to results generated by discriminant analysis, individual judgments, and group judgments. A partial intersection of predictive attributes occurs. More importantly, for all cases tested, the inductively produced knowledge structures perform better than the competing models. [ABSTRACT FROM AUTHOR]

This paper presents a heuristic procedure for balancing production assembly lines and a computer program for carrying out that procedure. This research was undertaken to investigate the application of complex information processing techniques (as used in producing the Chess Machine and Logic Theorist) to a typical industrial problem. The assembly line balancing problem is stated as: Given an assembly process made up of elemental tasks, each with a time required per unit of product and an ordering with other tasks, what is the least number of work stations needed to attain a desired production rate? The heuristic procedure for assembly line balancing consists of three phases: a) repeated simplification of the initial problem by grouping adjacent elemental tasks into compound tasks; b) solution of the simpler problems thus created by assigning tasks to work stations at the least complex level possible, breaking up the compound tasks into their elements only when necessary for a solution; c) smoothing the resulting balance by transferring tasks among work stations until the distribution of assigned time is as even as possible. The heuristics used in each phase are considered in some detail. Appropriate means for mechanizing such a procedure are discussed, and operating results of the program on actual problems are presented. [ABSTRACT FROM AUTHOR]

Our objective is to discuss the technique of digital system simulation. This procedure has already achieved a considerable stature in industrial and research organizations and promises to attain even greater importance in the future. Yet, with few exceptions (5), the published literature in the area consists of introductory expositions (10) or of descriptions of the solution of a particular problem (7) where the technique itself is correctly relegated to secondary position. The only publication devoted entirely to the topic (12) is a compendium of papers of the latter type with little attempt to summarize or generalize. Simulations are, of course, as varied as the systems which they represent but they do have certain common characteristics and problems. An identification of these problems would at least allow the investigator to anticipate them and plan accordingly. This paper consists of three sections. The first is an attempt to describe characteristics common to many system simulations; the second is a discussion of some problems involved in the construction of a digital simulator; and the third concerns problems that arise in the use of such a simulator. [ABSTRACT FROM AUTHOR]

Artificial intelligence (AI) is surpassing human performance in a growing number of domains. However, there is limited evidence of its economic effects. Using data from a digital platform, we study a key application of AI: machine translation. We find that the introduction of a new machine translation system has significantly increased international trade on this platform, increasing exports by 10.9%. Furthermore, heterogeneous treatment effects are consistent with a substantial reduction in translation costs. Our results provide causal evidence that language barriers significantly hinder trade and that AI has already begun to improve economic efficiency in at least one domain. This paper was accepted by Joshua Gans, business strategy. [ABSTRACT FROM AUTHOR]

Credit-risk evaluation is a very challenging and important management science problem in the domain of financial analysis. Many classification methods have been suggested in the literature to tackle this problem. Neural networks, especially, have received a lot of attention because of their universal approximation property. However, a major drawback associated with the use of neural networks for decision making is their lack of explanation capability. While they can achieve a high predictive accuracy rate, the reasoning behind how they reach their decisions is not readily available. In this paper, we present the results from analysing three real-life credit-risk data sets using neural network rule extraction techniques. Clarifying the neural network decisions by explanatory rules that capture the learned knowledge embedded in the networks can help the credit-risk manager in explaining why a particular applicant is classified as either bad or good. Furthermore, we also discuss how these rules can be visualized as a decision table in a compact and intuitive graphical format that facilitates easy consultation. It is concluded that neural network rule extraction and decision tables are powerful management tools that allow us to build advanced and user-friendly decision-support systems for credit-risk evaluation. [ABSTRACT FROM AUTHOR]

This paper introduces a neural-net approach to perform discriminant analysis in business research. A neural net represents a nonlinear discriminant function as a pattern of connections between its processing units. Using bank default data, the neural-net approach is compared with linear classifier, logistic regression, kNN, and ID3. Empirical results show that neural nets is a promising method of evaluating bank conditions in terms of predictive accuracy, adaptability, and robustness. Limitations of using neural nets as a general modeling tool are also discussed. [ABSTRACT FROM AUTHOR]

Technical developments in electronic communication, computing, and decision support, coupled with new interest on the part of organizations to improve meeting effectiveness, are spurring research in the area of group decision support systems (GDSS). A GDSS combines communication, computing, and decision support technologies to facilitate formulation and solution of unstructured problems by a group of people. This paper presents a conceptual overview of GDSS based on an information-exchange perspective of decision making. Three levels of systems are described, representing varying degrees of intervention into the decision process. Research on GDSS is conceived as evolving over time from the study of simple "shell" systems, consisting of menus of features available for selection by a group, to consideration of sophisticated rule-based systems that enable a group to pursue highly structured and novel decision paths. A multi-dimensional taxonomy of systems is proposed as an organizing framework for research in the area. Three environmental contingencies are identified as critical to GDSS design: group size, member proximity, and the task confronting the group. Potential impacts of GDSS on group processes and outcomes are discussed, and important constructs in need of study are identified. [ABSTRACT FROM AUTHOR]

The discipline of modeling has advanced only slowly compared to disciplines concerned with analyzing and solving models once they are brought into being. Structured Modeling is an attempt to redress this imbalance. Structured Modeling aims to provide a formal mathematical framework and computer-based environment for conceiving, representing, and manipulating a wide variety of models. The framework uses a hierarchically organized, partitioned, and attributed acyclic graph to represent the semantic as well as mathematical structure of a model. The computer-based environment is evolving via experimental prototypes that provide for ad hoc query, immediate expression evaluation, solving simultaneous systems, and optimization. If successful, Structured Modeling will enable model-based work to be done with greater productivity and acceptance by nonspecialists, will exploit important developments in small computers, and will cross-fertilize management science/operations research, artificial intelligence, database management, programming language design, and software engineering. This paper is an introduction and status report on a long term project. The presentation is based largely on examples; rigorous development and details are left to a series of technical reports. [ABSTRACT FROM AUTHOR]

Extracting sentiment from text is a hard semantic problem. We develop a methodology for extracting small investor sentiment from stock message boards. The algorithm comprises different classifier algorithms coupled together by a voting scheme. Accuracy levels are similar to widely used Bayes classifiers, but false positives are lower and sentiment accuracy higher. Time series and cross-sectional aggregation of message information improves the quality of the resultant sentiment index, particularly in the presence of slang and ambiguity. Empirical applications evidence a relationship with stock values--tech-sector postings are related to stock index levels, and to volumes and volatility. The algorithms may be used to assess the impact on investor opinion of management announcements, press releases, third-party news, and regulatory changes. [ABSTRACT FROM AUTHOR]

This research is part of a larger effort to build machine-based tools for developing scientific theories. In analogy with the research process in empirical research, we describe a logical cycle of theory development: (1) starting with an informal version of a theory, (2) then moving to its formal representation, (3) applying formal logic to investigate this representation, and (4) using the results as feedback for the update/revision of the original theory. A central aspect of the logical cycle is the detection of the (hidden) implications of a theory (called "partial deductive closure"). In this paper, we present an algorithm that performs the partial deductive closure for a relevant class of theorems, while filtering out trivial results. The algorithm is applied to an important organization theory, Organizational Ecology, and is shown to generate new theorems of interest. [ABSTRACT FROM AUTHOR]

In this paper a heuristic approach for the dynamic multilevel multiitem lotsizing problem in general product structures with multiple constrained resources and setup times is proposed. With the help of Lagrangean relaxation the capacitated multilevel multiitem lotsizing problem is decomposed into several uncapacitated single-item lotsizing problems. From the solutions of these single-item problems lower bounds on the minimum objective function value are derived. Upper bounds are generated by means of a heuristic finite scheduling procedure. The quality of the approach is tested with reference to various problem groups of differing sizes. [ABSTRACT FROM AUTHOR]

Researchers in artificial intelligence and decision analysis share a concern with the construction of formal models of human knowledge and expertise. Historically, however, their approaches to these problems have diverged. Members of these two communities have recently discovered common ground: a family of graphical models of decision theory known as influence diagrams or as belief networks. These models are equally attractive to theoreticians, decision modelers, and designers of knowledge-based systems. From a theoretical perspective, they combine graph theory, probability theory and decision theory. From an implementation perspective, they lead to powerful automated systems. Although many practicing decision analysts have already adopted influence diagrams as modeling and structuring tools, they may remain unaware of the theoretical work that has emerged from the artificial intelligence community. This paper surveys the first decade or so of this work. [ABSTRACT FROM AUTHOR]

A collaborative filtering system recommends to users products that similar users like. Collaborative filtering systems influence purchase decisions and hence have become targets of manipulation by unscrupulous vendors. We demonstrate that nearest neighbors algorithms, which are widely used in commercial systems, are highly susceptible to manipulation and introduce new collaborative filtering algorithms that are relatively robust. [ABSTRACT FROM AUTHOR]

Machine learning methods are powerful tools for data mining with large noisy databases and give researchers the opportunity to gain new insights into consumer behavior and to improve the performance of marketing operations. To model consumer responses to direct marketing, this study proposes Bayesian networks learned by evolutionary programming. Using a large direct marketing data set, we tested the endogeneity bias in the recency, frequency, monetary value (RFM) variables using the control function approach; compared the results of Bayesian networks with those of neural networks, classification and regression free (CART), and latent class regression; and applied a tenfold cross-validation. The results suggest that Bayesian networks have distinct advantages over the other methods in accuracy of prediction, transparency of procedures, interpretability of results, and explanatory insight. Our findings lend strong support to Bayesian networks as a robust tool for modeling consumer response and other marketing problems and for assisting management decision making. [ABSTRACT FROM AUTHOR]

Neural networks have been advocated as an alternative to traditional statistical forecasting methods. In the present experiment, time series forecasts produced by neural networks are compared with forecasts from six statistical time series methods generated in a major forecasting competition (Makridakis et al. 1982); the traditional method forecasts were estimated by experts in the particular technique. The neural networks were estimated using the same ground rule as the competition. Across monthly and quarterly time series, the neural networks did significantly better than traditional methods. As suggested by theory the neural networks were particularly effective for discontinuous time series. [ABSTRACT FROM AUTHOR]

A laboratory experiment was used to evaluate the effects of anonymity and evaluative tone on computer-mediated groups using a group decision support system to perform an idea-generation task. Evaluative tone was manipulated through a confederate group member who entered supportive or critical comments into the automated brainstorming system. Groups working anonymously and with a critical confederate produced the greatest number of original solutions and overall comments, yet average solution quality per item and average solution rarity were not different across conditions. Identified groups working with a supportive confederate were the most satisfied and had the highest levels of perceived effectiveness, but produced the fewest original solutions and overall comments. [ABSTRACT FROM AUTHOR]

A problem posed by O. L. Deutsch (1988) as the Artificial Intelligence Design Challenge for the 1987 American Institute of Aeronautics and Astronautics (AIAA) Conference on Guidance, Navigation & Control (Monterey, CA, August 17-19, 1987) is formulated as a zero-one linear program. We show that the associated (relaxed) linear programming problem can be solved in polynomial time despite an exponential size of the proposed constraint system in terms of the underlying parameter n of the number of cities considered, when a nonlinear constraint of the problem is either ignored or approximated by linearization. We describe a software system AIAA/ SOLVER that we have implemented to solve the problem to optimality under an apparently weak assumption about its stochastic cost structure using branch-and-cut. [ABSTRACT FROM AUTHOR]

A major complaint of people who use "decision-making" computer programs is that these programs merely provide a final decision, and fail to present the supporting argumentation, in terms the user can understand. This article presents an approach that makes computer programs more "human-like" by basing them on human decision making behavior. Decision making processes of student financial analysts are captured by asking them to think aloud during their evaluation. These verbal traces, called protocols, are analyzed at various levels of detail, resulting in specific models of the decision making processes involved, the strategies used, and the task-specific (financial) knowledge that is required to perform the task. The models and strategies are translated into executable computer programs. Extensive comparisons between human behavior and model simulation output are provided, assessing the extent that the computer program "thinks" and "talks" like a human decision maker. Although the model clearly suffers from "linguistic rigidity." It does appear to perform the evaluation in a similar manner as the human decision maker, examining the same information in the same order, making the same inferences, and reporting the same conclusions. [ABSTRACT FROM AUTHOR]