Ali Aouad

Improving access to agronomic information could enhance the livelihoods of hundreds of millions of smallholder farmers in low-and middle-income countries. Delivering timely, actionable advice remains a last-mile challenge. AI-powered chatbots offer a low-cost, scalable channel for personalized advice, and are increasingly being adopted by businesses and policymakers, yet there is a marked absence of causal evidence on their effects on farming practices. We conducted a pre-registered randomized controlled trial with 1,935 smallholder rice farmers in the Philippines and India over one production cycle. Treatment-arm farmers received an AI advisor on WhatsApp or Messenger, developed with local experts and promoting sustainable rice cultivation practices. Control-arm farmers received a similar conversational AI chatbot without agronomic content, isolating the added effect of agronomic information. AI advisory changed how farmers managed their farms. Deliberate practice changes increased by a fifth (+0.10 SD), and the quality of farming operations improved (+0.08 SD). Effects were concentrated on crop protection and problem response, which were also the most frequent topics in conversation logs. Treatment-arm farmers reported shifting toward more integrated management of pests, weeds, and diseases (+0.11 SD) and were more likely to adopt new practices in these areas (+0.10 SD) compared with control-arm farmers. Treatment also increased reported avoidance of crop-residue burning by 7 percentage points relative to control, but did not significantly affect the adoption of other targeted sustainable practices. We find no statistically significant effects on the adoption of technologies or inputs other than pesticides, or on costs or yields in that cycle. These findings suggest that AI advisory can improve management practices that farmers can adjust using resources at hand, particularly in response to time-sensitive problems, while on-demand information alone may not have transformative effects without more proactive use or access to complementary resources.

The design of online algorithms for matching markets and revenue management settings is usually bound by the assumption that the demand process is formed by a fixed-length sequence of queries with unknown types, each drawn independently. This notion of serial independence implies that the demand of each type, that is, the number of queries of a given type, has low variance and is approximately Poisson distributed. This paper proposes a nonparametric framework for modeling arrival sequences in online stochastic matching that departs from the serial independent assumption. We propose two models, Indep and Correl, that capture different forms of serial correlations by combining a nonparametric distribution for the demand with standard assumptions on the arrival patterns: adversarial or random order. The Indep model can capture arbitrary serial correlations within each customer type but assumes cross-sectional independence across types, whereas the Correl model captures common shocks across customer types. We demonstrate that fluid relaxations, which rely solely on demand expectations, have arbitrarily bad performance guarantees. In contrast, we develop new algorithms that achieve optimal (constant factor) performance guarantees in each model. Our mathematical analysis includes tighter linear programming (LP) relaxations that leverage distribution knowledge and a new lossless randomized LP rounding scheme for Indep. We test our new LP relaxations and rounding scheme in simulations on real and synthetic data and find that they consistently outperform well-established matching algorithms, especially on real data sequences that exhibit greater demand variance. This paper was accepted by Itai Ashlagi, revenue management and market analytics. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2024.05424 .

The Pandora’s Box Problem with Sequential Inspections This paper studies the fundamental problem of choosing among competing uncertain alternatives when the decision maker can invest resources to reduce uncertainty. The novelty of our work lies in examining settings in which uncertainty can be reduced either through a small investment or eliminated altogether through a more substantial one. The central question is how to navigate the selection process efficiently in order to achieve a good outcome. This type of problem arises, for example, in job search, when firms must choose among a large number of candidates whose fit for a position is uncertain and when uncertainty can be reduced through a small effort (e.g., an online or artificial intelligence–assisted interview) or a more substantial one (e.g., an in-person visit). We show that, whereas the problem is computationally challenging in general, there are two types of indices that can be computed efficiently for each alternative and that capture the attractiveness of a small or a large investment. These indices can guide the selection process, yielding optimal results in some cases and near-optimal results in all others.

A fundamental issue faced by experience providers—ranging from retail to culture—is displaying a collection of items for physical and digital interactions. The arrangement of the exhibits in different locations, which we call the layout, affects the visitors’ choices over time and space, thereby driving their engagement with the offered experience. In a collaboration with the Van Gogh Museum (Netherlands), we develop a predict-then-optimize framework to inform such operational decisions. First, we propose a sequential choice model, called pathway multinomial logit, that represents visitor activity as a sequence of conditional logit outcomes influenced by the layout. Estimation on large-scale visitor activity logs recorded on multimedia guides reveals that increase in spatial distances and search distances on the multimedia guide interface are strongly correlated with a reduction of transition propensity between artworks, while also uncovering relationships with artwork characteristics and contextual features. Counterintuitively, in response to more congestion, visitors may interact with more exhibits, including less prominent artworks. Our model predicts the next visitor transition with an out-of-sample accuracy of 63%. We test the predictive accuracy of our model against several benchmarks and modified layouts. Finally, we formulate the layout optimization problem, where the goal is to assign artworks to different locations to maximize the expected length of visitors’ paths. We establish a strong inapproximability result for this new optimization setting. Our simulations suggest that optimized layouts might lift visitor engagement by improving proximity and retention exerted by the layout. This paper was accepted by David Simchi-Levi, operations management. Funding: A. Aouad acknowledges the financial support provided by the UK Research and Innovation Engineering and Physical Sciences Research Council [Grant EP/Y003721/1]. V. Martínez-de-Albéniz acknowledges the financial support provided by the Agencia Estatal de Investigación from the Spanish Ministry of Science and Innovation [Project PID2020-116135GB-I00 MCIN/ AEI / 10.13039/501100011033]. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2022.02024 .

Motivated by the successes of deep learning, we propose a class of neural network–based discrete choice models, called RUMnets, inspired by the random utility maximization (RUM) framework. This model formulates the agents’ random utility function using a sample average approximation. We show that RUMnets sharply approximate the class of RUM discrete choice models: Any model derived from random utility maximization has choice probabilities that can be approximated arbitrarily closely by a RUMnet. Reciprocally, any RUMnet is consistent with the RUM principle. Our approach is closely related to ranking-based models and mixtures of multinomial logits proposed in previous literature in a more general contextual setting. We derive an upper bound on the generalization error of RUMnets fitted on choice data and provide theoretical insights on their ability to predict choices on new unseen data depending on critical parameters of the data set and architecture. The models are estimated by leveraging open-source libraries for training neural networks. We find that RUMnets are competitive against several choice modeling and machine learning methods in terms of predictive accuracy on two real-world data sets. We also conduct synthetic experiments that isolate the effects of each architecture component. This paper was accepted by George Shanthikumar, data science. Funding: A. Désir is supported by the Desmarais Fund for Research in AI from INSEAD. Supplemental Material: The online appendices and data files are available at https://doi.org/10.1287/mnsc.2023.02189 .

We introduce the click-based MNL choice model, a framework for capturing customer purchasing decisions in e-commerce settings. Specifically, we augment the classical Multinomial Logit choice model by assuming that customers only consider the items they have clicked on before they proceed to compare their random utilities. In this context, we study the resulting assortment optimization problem, where the objective is to select a subset of products, made available for purchase, to maximize the expected revenue. Our main algorithmic contribution comes in the form of a polynomial-time approximation scheme (PTAS) for this problem, showing that the optimal expected revenue can be efficiently approached within any degree of accuracy. To establish this result, we develop several technical ideas, including enumeration schemes and stochastic inequalities, which may be of broader interest. Using data from Alibaba’s online marketplace, we fit click-based MNL and latent class MNL models to historical sales and click data in a setting where the online platform recommends a personalized six-product display to each user. We propose an estimation methodology for the click-based MNL model that leverages clickstream data and machine learning classification algorithms. Our numerical results suggest that clickstream data are valuable for predicting choices and that the click-based MNL model can outperform standard logit-based models in certain settings. This paper was accepted by Chung Piaw Teo, optimization. Funding: D. Segev was supported by the Israel Science Foundation [Grant 1407/20]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/mnsc.2021.00281 .

The design of online policies for stochastic matching and revenue management settings is usually bound by the Bayesian prior that the demand process is formed by a fixed-length sequence of queries with unknown types, each drawn independently. This assumption of serial independence implies that the demand of each type, i.e., the number of queries of a given type, has low variance and is approximately Poisson-distributed. Thus, matching policies are often based on "fluid" LPs that only use the expectations of these distributions.