Online marketplaces will be transformed by autonomous AI agents acting on behalf of consumers. Rather than humans browsing and clicking, AI agents can parse webpages or interact through APIs to evaluate products, and transact. This raises a fundamental question: what do AI agents buy—and why? We develop ACES, a sandbox environment that pairs a platform-agnostic agent with a fully programmable mock marketplace to study this. We first explore aggregate choices, revealing that modal choices can differ across models, with AI agents sometimes concentrating on a few products, raising competition questions. We then analyze the current drivers of choices through randomized experiments on product positions and listing attributes. Models show sizeable and heterogeneous position effects: all favor the top row, yet different models prefer different columns, undermining the assumption of a universal ''top'' rank. They penalize sponsored tags, reward endorsements, and sensitivities to price, ratings, and reviews are directionally as expected, but vary sharply across models. Our findings reveal how AI agents behave in e-commerce, and surface concrete monitoring, seller strategy, platform design, and regulatory questions.

Emerging latency-critical applications such as cloud gaming and industrial automation demand agile and ultra-low-latency service delivery, which traditional network appliances struggle to support. Network Function Virtualization (NFV) addresses this by chaining Virtual Network Functions (VNFs) into Service Function Chains (SFCs). Parallelized SFCs (PSFCs) reduce service delay by executing independent VNFs concurrently, but introduce significant copy/merge and buffering overheads due to synchronization delays across branches. Moreover, dynamic PSFC arrival rates complicate efficient VNF placement decisions. This paper presents RL-PARETO, a hybrid deep reinforcement learning approach that adaptively orchestrates parallel VNFs while minimizing parallelization overheads and satisfying SLA constraints. RL-PARETO uses a graph transformer encoder with dual pointer-network heads to jointly generate PSFC partitions and VNF placements in a single pass. Training integrates Proximal Policy Optimization (PPO) for stable exploration with a Double-DQN critic for efficient value estimation. A fallback heuristic ensures feasible deployments under resource constraints. Extensive evaluations across diverse network topologies demonstrate that RL-PARETO achieves up to $10 \%$ higher acceptance rate and $15 \%$ reduction in merge buffer overhead, while maintaining robust performance under dynamic conditions.

Decision-making often requires an individual to navigate a multitude of options with incomplete knowledge of their own preferences. Information provisioning tools such as public rankings and personalized recommendations have become central to helping individuals make choices, yet their value proposition under different marketplace environments remains unexplored. This paper studies a stylized model to explore the impact of these tools in two marketplace settings: uncapacitated supply, where items can be selected by any number of agents, and capacitated supply, where each item is constrained to be matched to a single agent. We model the agents utility as a weighted combination of a common term which depends only on the item, reflecting the item's population-level quality, and an idiosyncratic term, which depends on the agent-item pair capturing individual-specific preferences. Public rankings reveal the common term, while personalized recommendations reveal both terms.

Recommendation systems are widespread, and through customized recommendations, promise to match users with options they will like. To that end, data on engagement is collected and used. Most recommendation systems are ranking-based, where they rank and recommend items based on their predicted engagement. However, the engagement signals are often only a crude proxy for user utility, as data on the latter is rarely collected or available. This paper explores the following question: By optimizing for measurable proxies, are recommendation systems at risk of significantly under-delivering on user utility? If that is indeed the case, how can one improve utility which is seldom measured? To study these questions, we introduce a model of repeated user consumption in which, at each interaction, users select between an outside option and the best option from a recommendation set. Our model accounts for user heterogeneity, with the majority preferring “popular” content, and a minority favoring “niche” content. The system initially lacks knowledge of individual user preferences but can learn these preferences through observations of users’ choices over time. Our theoretical and numerical analysis demonstrate that optimizing for engagement signals can lead to significant utility losses. Instead, we propose a utility-aware policy that initially recommends a mix of popular and niche content. We show that such a policy substantially improves utility despite not measuring it. As the platform becomes more forward-looking, our utility-aware policy achieves the best of both worlds: near-optimal user utility and near-optimal engagement simultaneously. Our study elucidates an important feature of recommendation systems; given the ability to suggest multiple items, one can perform significant exploration without incurring significant reductions in short term engagement. By recommending high-risk, high-reward items alongside popular items, systems can enhance discovery of high utility items without significantly affecting engagement.

Motivated by matching platforms that match agents in a centralized manner, we introduce a model of dynamic two-sided matching where both demand and supply are heterogeneous with many types and the pool of supply units is limited. We model heterogeneity on the two sides of the market by i.i.d. demand weight vectors and i.i.d. supply feature vectors, with possibly different distributions. The matching of a demand-supply pair generates a utility that depends on their weight and feature vectors. To reflect the realistic structure of a heterogeneous matching market while also avoid impossibility results, we consider various levels of assumptions (in particular, the spatial structure) on matching utilities and feature distributions. The goal of the centralized platform is to dynamically assign supply units to sequentially arriving demand units in order to maximize utility. Many popular heuristic policies are either sub-optimal (like the myopic policy) or computationally inefficient (like the certainty equivalent policy). We propose a forward-looking supply-aware policy dubbed Simulate-Optimize-Assign-Repeat (SOAR) that combines practicality and strong theoretical guarantee. Inspired by model predictive control (MPC), SOAR leverages the power of simulation to balance between producing immediate high match utility and preserving valuable supply for future demands. We use regret as our performance metric for matching policies, specifically the additive loss relative to the utility per match achievable in the continuum limit (n → ∞). Under mild regularity assumptions on the offline matching instances, we prove that SOAR achieves the optimal regret scaling (up to a log factor). We further characterize the optimal regret scaling for interesting classes of problems with additional model structure, in particular, two classes of utility functions: (i) the "spatial utilities", namely the negative p-th power of the Euclidean distance between the supply and demand vectors where p ≥ 1; and (ii) the dot-product utility (equivalently p = 2 of (i)), and two classes of distributions: (i) both supply and demand distributions are smooth (a more stringent assumption) and (ii) supply and demand distributions are supported over compact sets (a mild assumption). En route to proving our guarantees we develop a novel framework for analyzing the performance of our SOAR policy which may be of wider applicability and independent interest. As a corollary of our techniques, we also resolve an open problem posed in Kanoria 2022.

We study the multi-secretary problem with capacity to hire up to B out of T candidates, and values drawn i.i.d. from a distribution F on [0,1]. We investigate achievable regret performance, where the latter is defined as the difference between the performance of an oracle with perfect information of future types (values) and an online policy. While the case of distributions over a few discrete types is well understood, very little is known when there are many types, with the exception of the special case of a uniform distribution of types. In this work we consider a larger class of distributions which includes the few discrete types as a special case. We first establish the insufficiency of the common certainty equivalent heuristic for distributions with many types and "gaps" (intervals) of absent types; even for simple deviations from the uniform distribution, it leads to regret Θ(√T), as large as that of a non-adaptive algorithm. We introduce a new algorithmic principle which we call "conservativeness with respect to gaps" (CwG), and use it to design an algorithm that applies to any distribution. We establish that the proposed algorithm yields optimal regret scaling of ~Θ (T1/2 - 1/(2(β + 1))) for a broad class of distributions with gaps, where β quantifies the mass accumulation of types around gaps. We recover constant regret scaling for the special case of a bounded number of types (β=0 in this case). In most practical network revenue management problems, the number of types is large and the current certainty equivalent heuristics scale poorly with the number of types. The new algorithmic principle called Conservatism w.r.t Gaps (CwG) that we developed, can pave the way for progress on handling many types for the broader class of network revenue management problems like order fulfillment and online matching.

Recent work in networked systems has shown that using aerial imagery for farm monitoring can enable precision agriculture by lowering the cost and reducing the overhead of large scale sensor deployment. However, acquiring aerial imagery requires a drone, which has high capital and operational costs, often beyond the reach of farmers in the developing world. In this paper, we present TYE (Tethered eYE), an inexpensive platform for aerial imagery. It consists of a tethered helium balloon with a custom mount that can hold a smartphone (or a camera) with a battery pack. The balloon can be carried using a tether by a person or a vehicle. We incorporate various techniques to increase the operational time of the system, and to provide actionable insights even with unstable imagery. We develop path-planning algorithms and use that to develop an interactive mobile phone application that provides the user instant feedback to guide users to efficiently traverse large areas of land. We use computer vision algorithms to stitch orthomosaics by effectively countering wind-induced motion of the camera. We have used TYE for aerial imaging of agricultural land for over a year, and envision it as a low-cost aerial imaging platform for similar applications.

A single server with variable speed and a finite buffer is considered under a maximum packet drop probability constraint. The cost of processing by the server is a convex function of the speed of the server. If a packet arrives when the buffer is full, it is dropped instantaneously. Given the finite server buffer, the objective is to find the optimal dynamic server speed to minimize the overall cost subject to the maximum packet drop probability constraint. Finding the exact optimal solution is known to be hard, and hence algorithms with provable approximation bounds are considered. We show that if the buffer size is large enough, the proposed algorithm achieves the optimal performance. For arbitrary buffer sizes, constant approximation guarantees are derived for a large class of packet arrival distributions such as Bernoulli, Exponential, Poisson etc.