Industry Signals
Instance-Aware Parameter Prediction Cuts E-CVRP Routing Cost by 0.28% on Benchmark 🔗
Foundation: The Vehicle Routing ProblemFinding minimum-cost routes for a fleet to serve a set of customers. See full concept page. becomes the Electric Capacitated Vehicle Routing Problem (E-CVRP) when vehicles are electric and must respect both cargo capacity limits and battery energy constraints that vary with load and distance. Metaheuristics (algorithms that guide a combinatorial search using high-level strategies rather than proving optimality) are the practical solver of choice for E-CVRP at operational scale, since exact solvers cannot handle real fleet sizes within route-planning time windows. A single globally-tuned parameter configuration rarely performs well across the full range of instance structures a real fleet encounters.
Yinghao Qin et al. (CEC 2026) address this by predicting instance-specific parameters for the Bilevel Late Acceptance Hill Climbing solver before any routing runs. An offline tuning step derives best parameter settings for each training instance; a regression model maps instance features (demand density, energy profile, charging station layout) to those settings for unseen instances. On the World Congress on Computational Intelligence (WCCI) 2020 benchmark across eight held-out instances, instance-aware configuration reduces the objective by 0.28% versus a globally-tuned baseline.
CSIRO Combinatorial Optimisation Outperforms Integer Linear Programming Baseline on Multi-Condition Metabolic Network Reconstruction 🔗
Foundation: A Genome-Scale Metabolic Model (GEM) is a computational map of all biochemical reactions inside a cell — used to simulate how an organism grows under different nutrient conditions. Gap-filling is the process of adding reactions not directly confirmed by genomic data to make the model's predictions match experimental observations; without it, large portions of the model produce biologically impossible results. Traditional gap-filling solves one growth condition at a time using Integer Linear Programming (ILP), which works for single conditions but often leaves the model inconsistent when the same reaction set is tested across multiple media simultaneously.
Philip Kilby et al. at CSIRO, Australia, tackle simultaneous gap-filling across ten or more growth conditions by framing reaction selection from an 11,000-reaction universal database as a combinatorial optimisation problem, solved with a hybrid metaheuristic and continuous Linear Programme (LP) relaxation. Tested on three bacterial strains, the approach improves Kendall Tau rank correlation by 7.3 percentage points and reduces Root Mean Squared (RMS) Error by 13.3% over the standard ILP baseline, while running faster.
Research Papers
Linear Decision Tree Policies for Integer Linear Programs 🔗
Foundation: An Integer Linear Program (ILP) is an optimisation problem where decisions must be integers and the objective and constraints are linear. The standard approach is branch-and-boundExact MIP algorithm that recursively partitions the search space and prunes provably suboptimal subtrees. See full concept page.: a solver that recursively partitions the decision space and prunes subtrees proven suboptimal. This approach solves each new cost instance from scratch, even when the set of feasible solutions stays fixed. This paper asks whether that repeated work can be eliminated by a decision policy synthesised once offline and then queried at near-zero cost for any new cost vector.
Guyard et al. prove that an offline-synthesised linear decision tree can represent an optimal policy for a fixed-feasible-set ILP: the tree maps any queried cost vector to an optimal solution through a sequence of linear comparisons, and a polynomial-size policy always exists. A practical synthesis framework constructs policies within a specific subclass. In computational experiments, pre-built policies return optimal solutions orders of magnitude faster than classical branch-and-bound on repeated queries. The case study below details the formulation and structural difficulty.
arXiv:2605.02582 · Guyard et al., May 2026📋 Case Study below ↓Simulation-Based Optimisation of Batting Order and Bowling Plans in T20 Cricket 🔗
Foundation: Twenty20 (T20) cricket is the shortest professional format of the game, with 20 overs per innings, where two recurring team decisions (batting order and bowling plan) directly affect the match outcome yet are typically set by intuition rather than quantitative analysis. A Markov Decision Process (MDP) is a mathematical framework for sequential decisions under uncertainty where the optimal action at each step depends on the current state; calibrated on ball-by-ball historical data, it converts match-state information into tractable win-probability calculations. This paper builds that framework from 1,161 Indian Premier League (IPL) ball-by-ball records spanning 2008 to 2025.
The authors profile each player across three match phases (Powerplay, Middle, Death) using James-Stein shrinkage (a statistical technique that pulls individual player statistics toward the league average when sample sizes are small, reducing estimation error for players with sparse phase data). Win and defend probabilities are evaluated via Monte Carlo simulation over 50,000 innings trajectories. Applied to two 2026 IPL matches, the optimal batting order improves Mumbai Indians' win probability by 4.1 percentage points (52.4% to 56.5%).
Term of the Day
Bilevel Optimisation
"A hierarchy of subsystems, each of which is in turn hierarchic in structure, is the fundamental architecture of complexity." — Herbert A. Simon, The Architecture of Complexity, Proc. American Philosophical Society (1962)
Bilevel optimisation is a class of mathematical programme in which one optimisation problem is nested inside another as a constraint. A leader solves an outer problem, but their feasible set requires that an inner decision-maker (the follower) is simultaneously optimising their own objective in response to the leader's choice. The leader cannot pick any solution they like: they are restricted to those solutions where the follower has no incentive to deviate. This makes the problem structurally different from both classical single-level optimisation and two-stage stochastic programming, where the inner problem models uncertainty rather than a rational agent with a conflicting goal.
A concrete example
A road authority (the leader) sets tolls on two highway routes. The authority wants to maximise toll revenue. Commuters (the follower) choose the route that minimises their own travel cost, which includes the tolls the authority sets. The leader must commit to tolls before observing which routes commuters take.
Without modelling the follower: the authority sets tolls at £8 per trip and expects full revenue. But at £8, all commuters divert to the free surface street. Revenue: £0. The solution that looks best for the authority in isolation is not feasible as a bilevel solution because the follower would not comply with it.
With bilevel structure: the authority models the commuter routing decision exactly as a nested optimisation problem. The authority's outer problem searches over toll policies that, once commuters respond optimally, yield the highest revenue. The result is a lower toll that keeps the highway attractive to commuters while still generating positive revenue. This is the bilevel-optimal solution: the best the leader can achieve knowing the follower will react rationally.
Why practitioners misread this
The most common confusion is with two-stage stochastic programming, which also involves an inner problem. In two-stage programming, one decision-maker splits their decisions across time, and the inner problem represents recourse actions taken after uncertainty is revealed. The inner problem is not an agent with their own objective: it is a passive scenario drawn from a distribution. In bilevel, the follower is a rational optimiser. A commuter choosing the cheapest route is not a scenario; they will actively adjust their behaviour in response to the leader's choice, and that is exactly what makes the problem bilevel.
The second misread is confusing bilevel with min-max robust optimisation. In min-max, a single decision-maker chooses an action to perform as well as possible against the worst-case scenario. In bilevel, the follower minimises their own objective, which is not necessarily the leader's worst case. A commuter minimising travel cost may or may not produce the outcome worst for the toll authority, depending on the network structure. Bilevel does not assume adversarial intent; it assumes rational self-interest. The two coincide only in zero-sum games.
Where this shows up in practice
Bilevel optimisation appears wherever a leader must commit to a decision that then constrains a follower's rational response. Toll and tariff design: a regulator sets prices; users optimise their own consumption against those prices. Network interdiction: an attacker selects links to remove from a supply network; a defender re-routes optimally around the damage. Algorithm configuration: as in today's E-CVRP paper, the outer problem sets solver parameters; the inner problem is the solver running to completion on an instance. Supply chain contracting: a manufacturer sets wholesale prices; retailers maximise their own margins. The diagnostic question to ask when evaluating any hierarchical model: does the inner decision-maker have their own objective function, or are they a passive scenario drawn from a distribution? If the former, the problem is bilevel.
- E-CVRP LAHC Predicting solver parameters from instance features before any routing runs reduced the objective value by 0.28% on eight held-out benchmark instances, eliminating the gap between global and instance-optimal tuning without adding run-time overhead.
- GEM gap-filling (CSIRO) Framing metabolic network reconstruction as a single combinatorial optimisation problem solved across all growth conditions simultaneously improved ranking accuracy by 7.3 percentage points and cut prediction error by 13.3% over the standard ILP baseline.
- Linear Decision Trees Pre-building a linear decision tree policy over a fixed feasible set eliminates the branch-and-bound search for every new cost query; the pre-built policy returns optimal solutions orders of magnitude faster on repeated queries, at the cost of offline synthesis time.
- T20 batting order (IPL) Evaluating all feasible batting arrangements with a Markov chain calibrated on 1,161 ball-by-ball IPL records moved Mumbai Indians' win probability from 52.4% to 56.5%, a 4.1 percentage-point improvement over the coaches' intuitive order.