Research

When servers choose their own pace, managers must staff enough or pay enough. In a game-theoretic many-server queueing model, servers choose how fast to work in response to staffing, compensation, and admission decisions; loading underpaid servers past a tipping point triggers rebellion and a sharp drop in performance.

This paper pins down how the classical delay-probability formula behaves as service systems grow large. The Erlang-C formula, the probability that an arriving customer must wait, underpins staffing and optimization in many-server queueing systems. Its asymptotic behavior is characterized across the many-server limiting regimes, and the critically loaded case, where the literature had no limiting value, is settled through extensions of the square-root safety staffing rule.

Keeping every server busy is not always optimal: when utilization itself is costly, turning some arrivals away so servers can idle is asymptotically best. In a many-server queue with general arrivals, service times and patience, trading off abandonment and holding costs against the cost of server utilization, the fluid control problem shows that non-idling disciplines are not in general optimal, and an admission-control policy that guarantees servers sufficient idle time is asymptotically optimal.

When customers and servers are both strategic in the same many-server queue, the conclusions differ fundamentally from models where only one side is. Earlier work lets either customers decide whether to join or servers decide how hard to work, never both at once; this paper initiates the study of their joint strategic behavior, contrasting individually optimal with socially optimal outcomes.

When servers race for a winner-takes-all prize, adding a competitor can lower effort so much that the extra server adds no capacity. In a racing queue, strategic servers work on the same job and the first to finish earns a prize, as in cryptocurrency mining or bug bounties; equilibrium service rates are derived in closed form across overloaded, critically loaded and underloaded regimes, and effort can be non-monotone in server count, revealing an ideal level of competition. With the prize endogenized, the profit-maximizing prize sits systematically below the surplus-maximizing one, with relative surplus loss uniformly bounded.

Surge pricing and virtual queues are two ways a ride-hailing platform can ration peak-hour rides: raise prices, or make riders wait in a virtual queue. A queueing network model of a ridesharing platform compares the two mechanisms on efficiency and equity metrics and on consumer welfare.

Estimate the unknowns first, then schedule with a simple index rule: the gap to the full-information benchmark grows only logarithmically over time. In a multiclass many-server queue with abandonment and unknown service and patience rates, the Learn-then-Schedule policy estimates the primitives from observed data and then follows the simple, asymptotically optimal index rule, achieving the best possible regret rate.

A model-free reinforcement-learning strategy for placing orders on a limit order book, simple enough to run from a lookup table, beat benchmark strategies in a collaborating market-making firm's backtester. Q-learning with state aggregation, trained on event-by-event order-book data, produced a strategy that passed in-sample and out-of-sample testing and outperformed the benchmarks; the firm sought to put it into production.

On real data from St. Paul, Minnesota, a fairness-aware learned dispatch policy beats sending the nearest ambulance, gaining substantial fairness with minimal loss of efficiency. Emergency dispatching is formulated as a fairness-aware average-reward dynamic program, using α-fairness to weigh efficiency against equitable access across regions, and solved with a temporal-difference learning algorithm with guaranteed convergence. The learned policy beats nearest-unit dispatch on five efficiency and fairness metrics and stays robust to time-varying arrivals and non-Markovian service times.

When platforms learn policies from A/B tests, some of them contaminated, trusting only the experiments that agree with each other recovers the full gain an oracle would. A/B tests are natural instruments for levers set from observational logs, but bundled or overlapping treatments can move the reward directly; the method fits one value function per experiment and combines the largest group whose fitted functions agree. With four of ten simulated experiments contaminated, it captures the full oracle gain, where pooled instrumental-variable Q-learning captures 10%.

Continuity with one's own physician competes with fast access; a sequence of queueing models shows how to protect continuity for those who benefit, pool access for those who do not, and empower patients to choose. Primary care faces a trade-off between relational continuity, seeing the same physician over time, and access, being seen quickly; models of increasing sophistication identify which patients gain from continuity and how the choice between the two should be organized.

Two shortcomings of the elective caesarean pathway motivated this collaboration with Cambridge University Hospitals: patients known early could wait weeks or months for a confirmed date although the hospital already knew they needed planned surgery, while patients revealed later needed access to a limited elective list after much of the demand had accumulated. These are different service problems, one about the timing and reliability of information given to patients, the other about access to planned capacity; a reservation policy addresses both with explicit case-loss trade-offs. The resulting tool, which sets how much capacity can be committed early while retaining enough for later-revealed demand, is being piloted at Cambridge University Hospitals.

In a national cohort of 1,027,707 first-time mothers in NHS maternity care in England, 2021–2025, each one-point rise on a 0–10 neighborhood deprivation score carried 3.4% higher odds of early-onset pre-eclampsia (adjusted odds ratio 1.034). Adjusting for theorized mediators attenuated this modestly (1.023) and for hospital site further (1.016), so both individual risk factors and maternity-site variation contribute. Associations were similar but stronger among 940,505 multiparous women, and weak or absent for late-onset pre-eclampsia.

Childcare waitlists are long while slots sit empty because stale 'phantom' listings clog them. Families who have found care elsewhere remain listed but have gone phantom. Two remedies combat this phantom paradox: verification clears phantom listings before offers are made, and mixed-age classrooms let older-child slots serve younger children; both raise utilization and shorten waits.

Across 11,271 LinkedIn posts, the same AI-risk disclosure signal tracks sharply different financial patterns depending on a firm's place in the AI value chain: disclosure is not one-size-fits-all. An auditable, agentic measurement system extracts the signal and relates it to firms' positions in the AI value chain and their business networks, giving managers and regulators a practical early-warning system and a clear reason not to impose uniform disclosure expectations.

Nudge emails on a mental-health platform had no average effect on follow-up bookings; exploratory analysis finds a simple prompt helps clients early in therapy. In a pre-registered randomized field experiment on Safe Space, a Southeast Asian platform, three nudge emails (a basic prompt, a social-information nudge, and a research-based nudge) were broadcast to eligible clients, and none moved bookings on average. Among clients early in therapy the basic nudge beats richer content, so targeting simple nudges early may generate more bookings with fewer emails.