The deep end of the toolkit.
Each card below is a methodological capability we have led, co-authored, or supervised. Some has been published in a peer-reviewed journal, others written up as a technical report. Most improved on the prior state of the art by orders of magnitude. When the decision has to be provably good, this is the research bench we draw on.
Exact Choice-Based Pricing, Up to 1.67 Million Times Faster Than the MILP Baseline
We developed three exact algorithms for continuous pricing under any advanced discrete choice model (mixed logit, nested logit, probit, anything with linear-in-price utility): a spatial Branch-and-Bound, a Branch-and-Benders Decomposition variant, and a polynomial-time Breakpoint Exact Algorithm. On the standard parking-pricing benchmark with mixed-logit demand, the BEA solves 1,000,000 demand scenarios in 77 seconds — a 1.67-million-× speedup over Gurobi's MILP and a 300× speedup over CoBiT, the previous ML-specific state-of-the-art. Published in OR Spectrum.
A Pricing Heuristic With <0.2% Optimality Gap and Up to 37 Million × Speedup Over the State-of-the-Art
We introduced the Breakpoint Heuristic Algorithm and its Iterated Local Search extension: a fast, scalable heuristic for the capacitated choice-based pricing problem that delivers sub-0.2% average optimality gap on benchmarks where every exact method times out. Versus the previous capacitated state-of-the-art (LAG), it's at least 79× faster. Versus the previous mixed-logit-specific state-of-the-art (CoBiT), it's up to 37 million × faster — while the ILS extension matches every known optimum. Published open-access in Computers & Operations Research.
BHAMSLE: +17.2% Log-Likelihood Lift on London Mode-Choice Mixture, in Half the Runtime of Multistart Biogeme
We took the Breakpoint Heuristic Algorithm developed in our OR Spectrum pricing work and adapted it to Maximum Simulated Likelihood Estimation of advanced discrete-choice models with latent classes and continuous mixtures. On the London mode-choice discrete-continuous mixture, BHAMSLE-initialized Biogeme reaches a 17.2% better log-likelihood than default initialization (LL = −243.68 vs −294.38), in roughly half the wall-clock time that Multistart Biogeme with 20 restarts needs to match the same quality. Submitted to Elsevier.
A Decentralized Facility-Location Optimizer That Solves 11× More Scenarios Than the Textbook MILP
We supervised the design of a decentralized capacitated facility-location optimizer that integrates advanced discrete-choice demand directly into the MILP, with capacity handled by exogenous priority queues. On the real Dresden Gymnasium-location dataset, the Progressive Scenario Expansion method solves instances of 14 schools × 558 students × 110 scenarios within a 24-hour limit — where the textbook MILP stalls beyond 10 scenarios. With the full feature stack (priority valid inequalities + complete warm start + extended-entropy ordering), PSE beats its basic configuration by up to 95% and the RAW MILP by 28% on cases both can solve.
Electric Autonomous Ride-Sharing Optimization, 31× Faster Than the Published Branch-and-Price Baseline With Up to 11× Tighter Gaps
We supervised the design of an event-based mixed-integer formulation for the electric autonomous Dial-A-Ride Problem that beats the published Branch-and-Price baseline of Su (2023) by 31× on average runtime and 11× on average optimality gap on hard instances. The companion Adaptive Large Neighborhood Search heuristic provides a deployment lane: 10× faster than the exact method at less than 1% loss in objective value. On the largest benchmark instance (r8-96, 96 users, 8 vehicles), ALNS finds a better solution than the exact method in 9× less time.
Joint Mode-and-Destination Choice Sets, Now Live in EPFL's OASIS Pipeline
We co-authored the first choice-set generator that handles the mode/destination coupling jointly and ships a closed-form bias-correction term any activity-based-model estimator can plug into. Now the choice-set-generation building block of the OASIS activity-based modeling pipeline at EPFL Transp-OR.
Constraint Programming, 750× Faster Than MILP on Activity-Based Scheduling
We co-supervised a head-to-head benchmark of constraint programming versus the dominant MILP formulation for activity-based scheduling, and built CP variants that solve the problem 750× faster on 5-activity instances. The deeper lesson: not every optimization problem deserves the same hammer.