My Name Is Earl and Large Neighbourhood Search: Sometimes You Have to Break the Plan
An intuitive exploration of Large Neighbourhood Search, using My Name Is Earl and Yorkshire logistics examples.
6 min read →By Richard Faint · 15 August 2026 · 4 min read
TL;DR
There is no single best optimisation algorithm. Real routing systems combine decomposition, construction heuristics, local search and metaheuristics because each solves a different part of the problem. Earl never needed a perfect way to finish his list; he needed a good way to decide what to do next.
Earl Never Got His Finale
When this series began, using My Name Is Earl to explain route optimisation seemed like a slightly ridiculous premise, which is exactly why I liked it! Earl Hickey has a list, a clear objective and an apparently linear strategy i.e. put right the bad things he has done until everything is crossed off.
Except Earl never finished the list, the final episode ended with a “To Be Continued…” cliffhanger before the programme was cancelled, which makes it an even better analogy for operational optimisation (and one of those awful network decisions - don’t get me started on Firefly and Angel!). Real systems rarely reach a neat final state either orders change, traffic builds and something that looked solved an hour ago becomes a problem again. The lesson of the series therefore became less about individual algorithms and more about knowing what kind of search a particular problem needs.
Make the Problem Smaller, Then Build Something Useful
We started with grouping because optimisation often begins before a route optimiser sees a route. K-means, DBSCAN, hierarchical clustering, geographic zoning and sweep algorithms all show that a large problem often becomes manageable only after it has been divided into sensible pieces. A brilliant routing algorithm cannot rescue a fundamentally poor grouping decision.
Clarke-Wright Savings then gave us a sensible first solution.
Randy Starts Rearranging the Route
Local search, 2-Opt and 3-Opt showed how much improvement can come from small changes e.g. reversing part of a route, reconnecting two edges, or swapping two visits. Eventually, however, every nearby move can look worse even though much better solutions exist elsewhere. Simulated Annealing escapes by occasionally accepting deterioration, while Tabu Search uses memory to stop the optimiser returning immediately to decisions it has just abandoned. One tolerates a temporary mistake; the other remembers previous ones.
Ants, Genetics and Breaking the Plan
Ant Colony Optimisation allowed many agents to explore alternatives and reinforce promising choices, while Genetic Algorithms maintained competing solutions and created new candidates through selection, crossover and mutation.
Large Neighbourhood Search then challenged the idea that improvement must always be incremental. If small changes stop working, remove part of the solution and rebuild it. Sometimes refinement is simply polishing the wrong answer.
So Which Optimisation Algorithm Won?
There is no universally best optimisation algorithm because there is no universally identical optimisation problem. A twenty-stop delivery route is not the same as a national network containing thousands of jobs, multiple depots, heterogeneous vehicles, time windows and loading constraints. Even “best” is difficult to define because real transport operations balance mileage, driving time, empty running, vehicle fill, overtime, service levels and plan stability, with improvements in one measure often making another worse.
That is why serious optimisation systems are usually hybrids. Geographic grouping reduces the search space, a construction heuristic produces a baseline, local search cleans it up, Simulated Annealing or Tabu Search escapes local optima, and Large Neighbourhood Search restructures the plan when incremental improvement has run out of road.
The useful question is not “Which optimisation algorithm is best?” but “Which combination of decomposition, representation, heuristics, neighbourhoods, constraints and search strategies produces the best decisions for this operation?”
Earl Finishes the List, Sort Of
Greg Garcia later explained that his intended ending was not for Earl simply to cross off one final item. Earl would struggle with a difficult item, discover that someone else had created a list after being inspired by him, and realise that the idea had spread beyond his own objective and was making the world a much better place.
That feels like the right ending for this series. Operational optimisation is never really finished. Today’s plan is the best decision we can justify using today’s orders, vehicles, drivers and constraints; tomorrow the system will be different and the search begins again.
Earl never needed a perfect list or a perfect algorithm, he needed a practical way of deciding what to do next.
In the end, that is what optimisation is for.