In short: An abbreviation in planning almost always names one of three things. A metric, which measures a plan or a forecast. A problem type, meaning the rules your plan must keep. Or a method, which finds the plan. Once you know which of the three you’re looking at, you’re halfway there.
Why this page
Today, a software demo often ends with a slide like this: “MILP-based, CP-SAT for sequencing, WAPE down on the A items.” Everyone nods. After the meeting you look up three of the terms. You find three explanations, each written for mathematicians. And the question that matters stays open: what does that mean for our operation?
This page is the short version. Every abbreviation we use on this site gets two to four sentences in plain words. Each entry also says what the term hides, or when it misleads you. You also see where it comes from: we link the original paper or the documentation. Where one of our pages uses a term, we link to it, so you can see the term at work.
The terms come in six groups. Forecast metrics are first, then stock and systems, production and routing. The methods and the solvers come last. Jump in from the contents, or search the page with Ctrl+F (Cmd+F on a Mac). Every entry has its own link, so you can send a colleague straight to the one they need.
A to Z
- ABC
- ADI
- ADIDA
- API
- APS
- Branch and bound
- CBC
- CP
- CP-SAT
- CPLEX
- Croston’s method
- CRP
- CV²
- CVRP
- Cycle service level and fill rate
- DC
- EOQ
- ERP
- EV
- FJSP
- FMCG
- Forecast bias
- FSP
- FVA
- GHG
- GRG
- Gurobi
- Heuristic and metaheuristic
- HFS
- HiGHS
- IIS
- JSP
- LMS
- LNS
- LP
- LP relaxation
- LPL
- LTL
- Makespan
- MAPE
- MDVRP
- MILP
- MOQ
- MPS
- MRO
- MRP
- NEH
- NP-hard
- Objective function and constraints
- OEE
- Optimality gap (MIP gap)
- OR
- OR-Tools
- Pareto front
- PDPTW
- PuLP
- PVRP
- RCCP
- RCPSP
- RFP
- SBA
- SCIP
- SDST
- SKU
- SLA
- SMED
- Solver
- Stochastic optimisation
- TCO
- TMS
- TOC
- TSB
- VBA
- VRP
- VRPTW
- VSS
- WAPE
- WIP
- WMAPE
- WMS
Forecast metrics
How you measure a forecast decides which forecast wins. These are the numbers we look at first, and what each one hides.
MAPE Mean absolute percentage error
For every period you take the error as a share of actual sales, then average those shares. It’s easy to explain and easy to break. A slow seller that sold 2 and was forecast at 4 counts as a 100% miss, and a period with zero sales can’t be scored at all. So MAPE punishes your slow movers and says little about the money at stake.
WAPE Weighted absolute percentage error
You add up all absolute errors and divide by total actual sales. Big sellers now weigh more than slow ones, and periods with zero sales no longer break the number. For most stock decisions this is the better headline figure. Report it next to the bias, because WAPE alone can’t tell you whether you miss high or low.
WMAPE Weighted mean absolute percentage error
Often just another name for WAPE. Some teams let an error on an expensive item count more. They weight by revenue or by margin, the profit per unit, instead of by units. Before you compare two reports, ask which weight each one uses. Two WMAPE figures with different weights can’t be compared.
Source
- Kolassa, S., Schütz, W. (2007). Advantages of the MAD/MEAN ratio over the MAPE. Foresight: The International Journal of Applied Forecasting 6.
At work on: Consumer goods and FMCG
Forecast bias
The average signed error: do you forecast too high or too low, on balance? Errors in both directions cancel out here, so a bias near zero with a large WAPE means you miss randomly. A steady positive bias fills your warehouse. A steady negative one empties your shelves.
At work on: Consumer goods and FMCG
FVA Forecast value added
FVA asks whether each step in your forecasting process makes the forecast better. You compare the error of three forecasts. The first is the naive forecast: it simply repeats last period’s sales. The second is the statistical forecast from your system. The third is the forecast after you’ve adjusted numbers by hand. All three must look equally far ahead, say four weeks. If the meeting that adjusts the numbers makes them worse, you can stop holding it. It’s the cheapest improvement most teams can make.
Source
- Gilliland, M. (2010). The Business Forecasting Deal. Wiley.
At work on: Consumer goods and FMCG
ADI Average demand interval
The average number of periods from one demand to the next. A part used in 6 of 24 months has an ADI of 4. Above an ADI of about 1.32, demand counts as intermittent, and classic smoothing methods start to forecast badly. It’s the first number to check for spare parts.
At work on: Spare parts and MRO
CV² Squared coefficient of variation
CV² measures how much order sizes vary when demand does occur. Together with the ADI it sorts your items into four groups: smooth, intermittent, erratic and lumpy. Each group gets its own forecasting method. The usual cut-offs, an ADI of 1.32 and a CV² of 0.49, come from Syntetos, Boylan and Croston (2005).
At work on: Spare parts and MRO
Croston’s method
A forecast built for parts with many zero periods. It forecasts two things separately: the size of a demand and the gap between demands. It only updates when a demand happens, so long quiet stretches don’t drag the forecast to zero. Its known flaw is that it forecasts a little high.
At work on: Inventory optimisation , Spare parts and MRO
SBA Syntetos–Boylan approximation
A corrected version of Croston’s method, named after the two researchers who found the fix. Croston’s method is a forecast for parts that are rarely needed. It always comes out a little too high. So SBA multiplies the Croston forecast by a number just below 1. That makes the forecast a little smaller, and it’s no longer too high. For most spare parts it’s a sensible default.
At work on: Spare parts and MRO
TSB Teunter–Syntetos–Babai method
Another Croston variant, named after its three authors. It updates the chance of a demand in every period, including the empty ones. When a part stops selling, the forecast falls with it. Use it where parts may go obsolete.
At work on: Spare parts and MRO
ADIDA Aggregate–disaggregate intermittent demand approach
You add up demand over longer stretches of time, say months into quarters. That leaves hardly any stretch with no demand at all. You forecast at that level with an ordinary method, then split the result back into months. It’s simple and often surprisingly robust for sporadic items.
At work on: Spare parts and MRO
Stock, supply chain and systems
The words you hear in planning meetings, and the names of the systems you plan with. Most of them name a system or a rule you already live with.
SKU Stock keeping unit
One item you count and plan on its own. The same T-shirt in three sizes is three SKUs. In many models the same item at two warehouses is two SKUs as well, because each location needs its own stock decision.
At work on: Inventory optimisation , Replenishment planning , Warehouse slotting and layout , Assortment and SKU rationalisation
ABC analysis
You sort items by their share of turnover or value. A few A items usually bring in most of the turnover. The many C items together bring in little. For each class you set two things: the service level, meaning how reliably the item must be in stock, and how often you check the stock. ABC analysis is a sensible first sorting. But it doesn’t show how hard an item is to forecast. So add an XYZ analysis, which sorts items by how much their demand varies.
Sources
- Dickie, H. F. (1951). ABC inventory analysis shoots for dollars, not pennies. Factory Management and Maintenance 109(7). The first description of ABC analysis.
- Silver, E. A., Pyke, D. F., Thomas, D. J. (2016). Inventory and Production Management in Supply Chains (4th ed.). CRC Press.
At work on: Inventory optimisation , Warehouse slotting and layout , Warehousing and distribution
Cycle service level and fill rate
Two different promises hide behind “service level”. Cycle service level is the chance that you don’t run out between two orders. Fill rate is the share of demand you serve straight from stock. A 95% cycle service level can still leave many orders short, so agree which one you mean before you set targets.
At work on: Field technician routing , Workforce shift scheduling , Inventory optimisation , Safety stock and reorder point
EOQ Economic order quantity
The order size that balances ordering cost against holding cost. Order too often and the fixed costs per order pile up. Order too much and the stock ties up cash. You work out the EOQ with a fixed formula. That formula assumes you sell the same amount every week and that your warehouse has unlimited space. So treat the result as a starting point, not a plan.
At work on: Inventory optimisation , Replenishment planning
MOQ Minimum order quantity
The smallest amount a supplier will sell you in one order. Even if you need less, you have to order at least this amount. A good model always keeps the minimum order quantity. It also shows you how much more you pay because of it, so you know when it’s worth negotiating.
At work on: Replenishment planning
ERP Enterprise resource planning
The system that holds your orders, stock and finance, such as SAP. It also holds the bills of materials: the lists of which parts go into which product. It records what happens and runs standard planning rules. It rarely weighs thousands of options against each other. That’s where an optimisation model comes in: it reads from your ERP and writes its plan back.
At work on: Production scheduling , Job shop scheduling , Inventory optimisation , Replenishment planning
MRP Material requirements planning
The function in your ERP that turns demand into purchase and production orders. It breaks each product down into its parts. Then it counts back from the delivery date: when do you have to order or make each part so it’s there in time? It assumes every machine has unlimited free hours. That’s why MRP plans often look fine on screen and fail on the shop floor.
Sources
- Orlicky, J. (1975). Material Requirements Planning. McGraw-Hill.
- Jacobs, F. R., Weston, F. C. (2007). Enterprise resource planning (ERP): A brief history. Journal of Operations Management 25(2).
At work on: Job shop scheduling , Inventory optimisation , ERP not good enough
APS Advanced planning and scheduling
Planning software that only schedules as much as your machines and people can handle. It takes the orders from your ERP and gives the shop floor its plan. Most APS tools use fixed rules and heuristics, which are rules of thumb that find a usable plan fast. That works until your plant has a rule the tool wasn’t built for.
At work on: Legacy system replacement , Production scheduling , Manufacturing
RCCP Rough-cut capacity planning
Rough-cut capacity planning starts from planned sales, not from single orders. It groups similar products into product families. Then it checks whether your busiest machines and teams have enough hours for them over the coming months. It’s fast and coarse. You use it to decide early whether you need extra shifts or overtime.
Source
- Vollmann, T. E., Berry, W. L., Whybark, D. C., Jacobs, F. R. (2005). Manufacturing Planning and Control for Supply Chain Management (5th ed.). McGraw-Hill.
At work on: Machine capacity planning
CRP Capacity requirements planning
Capacity requirements planning starts from each single order. Every order has a list of work steps that says which machine each step needs. CRP adds up those hours per machine and week. It’s detailed and slow. It shows you in which week a machine would need more hours than it has. Don’t confuse CRP with rough-cut capacity planning (RCCP): they answer different questions.
Source
- Vollmann, T. E., Berry, W. L., Whybark, D. C., Jacobs, F. R. (2005). Manufacturing Planning and Control for Supply Chain Management (5th ed.). McGraw-Hill.
At work on: Machine capacity planning
WMS Warehouse management system
The system that runs a warehouse. It knows every storage place and the stock on it. And it hands out the work: which goods go onto the shelf and which are picked from the shelf for an order. The WMS carries out the work. It doesn’t optimise. An optimisation model can also decide which item goes in which place. Or which orders your staff pick together on one walk through the warehouse.
At work on: Live operations model
TMS Transport management system
The system for shipments, carriers, freight costs and tracking. Many include basic route planning. When your rules outgrow it, a routing model plans the tours and hands them back to the TMS for execution.
At work on: Real-time dispatch , Live operations model
DC Distribution centre
A warehouse whose job is to move goods on, not to store them for long. Network design decides how your sites are spread out. It asks: how many DCs do you need? Where? And which customers does each DC serve?
At work on: Supply chain network design , Warehouse slotting and layout , Network simulation and what-if
MRO Maintenance, repair and operations
The spare parts and supplies that keep plants and fleets running, as opposed to the materials that go into your product. Demand is sporadic and a missing part can stop a production line, which makes MRO stock planning its own discipline.
At work on: Inventory optimisation , Aviation and aerospace , Spare parts and MRO
FMCG Fast-moving consumer goods
Everyday products that sell quickly and cheaply: food, drinks, cleaning products, toiletries. You earn little on each unit, promotions are frequent and shelf life is short. Order too much and you throw goods away. Order too little and the shelf stands empty. That’s why a good forecast and timely restocking matter so much here.
At work on: Markdown and pricing optimisation , Consumer goods and FMCG
SLA Service level agreement
A promise to a customer with a number on it: on site within four hours, delivered by 10 am. A model can handle an SLA in two ways. Either every plan must always keep it. Or a late arrival is allowed, and the model adds a penalty cost for each one. Which one you pick changes the plan, so it’s worth deciding on purpose.
At work on: Field technician routing , Workforce shift scheduling , Field service management , Security and facility services
TCO Total cost of ownership
Price plus everything that comes after it: freight, duties, quality problems, stock you hold because of long lead times. Two suppliers with the same unit price can be far apart on TCO. It’s the number to compare bids on.
At work on: Procurement and sourcing
RFP Request for proposal
A formal invitation to suppliers to bid on a contract. With many transport routes or items, there are a great many ways to combine the bids. Finding the best combination is itself an optimisation problem.
At work on: Bid and RFP analysis
LMS Learning management system
Software that holds courses, enrolments and completion records. It knows who must attend what. It doesn’t decide when sessions run, in which room, with which trainer. That’s a scheduling problem on top.
At work on: Training course scheduling
Production and scheduling
Shop-floor measures and the names of the classic scheduling problems. The name tells you which model fits your plant.
OEE Overall equipment effectiveness
How much planned machine time turns into good parts. It multiplies three shares: availability (did it run?), performance (at full speed?) and quality (good parts?). A better schedule mostly moves availability, by cutting changeovers and idle time.
Source
- Nakajima, S. (1988). Introduction to TPM: Total Productive Maintenance. Productivity Press.
At work on: Production scheduling , Machine capacity planning , Manufacturing
SMED Single-minute exchange of die
A shop-floor method to shorten changeovers. You do as much of the setup work as you can while the machine is still running. SMED makes each changeover shorter. Sequencing decides how many changeovers you need. The two add up, so you don’t have to pick one.
Source
- Shingo, S. (1985). A Revolution in Manufacturing: The SMED System. Productivity Press.
At work on: Batch scheduling and changeover minimisation
SDST Sequence-dependent setup times
The changeover time depends on what ran before. Switching to a similar part takes minutes, switching to one that needs other tools can take an hour. With SDST the job order decides how much time you lose, so sequencing becomes the main lever.
At work on: Production scheduling , Job shop scheduling , Batch scheduling and changeover minimisation
WIP Work in progress
Jobs that have started but aren’t finished. Money is tied up in these half-finished jobs, and they take up space. The more of them wait in front of the machines, the longer each job takes from start to finish. A schedule that starts everything early looks busy and usually just piles up WIP.
Sources
- Little, J. D. C. (1961). A proof for the queuing formula L = λW. Operations Research 9(3). The law that links WIP, throughput and lead time.
- Hopp, W. J., Spearman, M. L. (2008). Factory Physics (3rd ed.). McGraw-Hill.
At work on: Production scheduling
TOC Theory of Constraints
A management method set out by Eliyahu Goldratt in 1984. Your plant never delivers more than its slowest step can handle. That step is the bottleneck. So you plan the bottleneck first and make sure it never runs out of work. In optimisation, the bottleneck is called the binding constraint: the one rule that limits your plan right now. A model can tell you which machine or rule that is today.
Source
- Goldratt, E. M., Cox, J. (1984). The Goal: A Process of Ongoing Improvement. North River Press.
At work on: Capacity planning and bottleneck analysis
Makespan
The time from the first job starting to the last job finishing. Minimising makespan packs the schedule tight. It’s a common target, but not always the right one: if due dates matter more, measure lateness instead.
At work on: Production scheduling , Job shop scheduling , Flow shop scheduling , Flexible job shop scheduling
JSP Job shop scheduling problem
Every job has its own route through the machines. One job goes to the lathe first, then the mill, then the drill. Another starts at the drill and ends at the lathe. The plan decides in which order each machine works through its jobs. Finding a good plan for this is hard. That’s why so many job shops plan on gut feel.
At work on: Job shop scheduling
FJSP Flexible job shop scheduling problem
A job shop where each step can run on one of several machines. The plan now decides two things: which machine and in what order. That’s closer to how most plants work. It also means many more possible plans, so it takes longer to solve.
Source
At work on: Flexible job shop scheduling
FSP Flow shop scheduling problem
All jobs pass the machines in the same order, like a production line. The simplest form is the permutation flow shop (PFSP). There, every machine works through the jobs in the same order, so no job overtakes another. The plan is then one single sequence of jobs. Fast rules of thumb such as the NEH heuristic find good plans for it.
At work on: Flow shop scheduling
HFS Hybrid flow shop
A flow shop where at least one stage has several parallel machines. Also called a flexible flow shop. Most production lines look like this: two mixing tanks, three filling lines, one packing station.
At work on: Flow shop scheduling
NEH heuristic
A rule of thumb for flow shops, named after Nawaz, Enscore and Ham. It sorts jobs by total processing time and inserts them one by one at the best spot in the sequence. It doesn’t prove the result is best, but it gives a strong plan in moments and a good start for better methods.
At work on: Flow shop scheduling
RCPSP Resource-constrained project scheduling problem
A project is made of tasks in a fixed order: B can’t start before A is done. Several tasks need the same crews, machines or budget, and there isn’t enough for all of them at the same time. You use this model to plan building sites, planned maintenance shutdowns and development projects. A Gantt chart, a bar chart with one bar per task, shows the result. The model works it out.
At work on: Project and task scheduling , Construction and project-based
Routing
Every routing problem is a VRP with letters added. Each letter is one more rule your routes must keep.
VRP Vehicle routing problem
Which vehicle visits which stops, in which order, so the total cost is lowest. It’s the base of every routing model. The plain VRP has no other rules. In practice, more rules almost always come on top, and each one adds a letter to the name. CVRP, for example, has a C for the capacity of the vehicles.
Source
At work on: Vehicle routing optimisation , Real-time dispatch , Logistics and transport
CVRP Capacitated vehicle routing problem
The VRP with a load limit per vehicle, in weight, volume, pallets or all three. It’s the first rule almost everyone needs.
Source
At work on: Vehicle routing optimisation
VRPTW Vehicle routing problem with time windows
Each stop must be served inside a window, say 8 to 10 am. Windows cut the number of possible routes sharply and make good routes much harder to find by hand. With soft windows a vehicle may arrive late, and the model adds a penalty cost for it. With hard windows no late arrival is allowed.
At work on: Vehicle routing optimisation , Last-mile route planning , Last-mile delivery
PDPTW Pickup and delivery problem with time windows
Each order has a pickup and a drop-off, and the same vehicle must do both, pickup first. Couriers, patient transport and on-demand services work like this. Because pickup and drop-off belong together, this problem is harder to solve than routing with time windows alone (VRPTW).
At work on: Pickup and delivery optimisation , Outgrown routing software
MDVRP Multi-depot vehicle routing problem
Vehicles start from several depots, and the model also decides which depot serves which customer. Some customers are about equally far from two depots. If you plan each depot on its own, each of these customers stays with the depot it has today. You never check whether the other depot could serve them for less. These customers are usually where you save the most.
At work on: Vehicle routing optimisation , Multi-depot routing
PVRP Periodic vehicle routing problem
Visits repeat on a cycle: twice a week, every second Monday. The model picks the visit days for each customer and the routes for each day together. Service routes, waste collection and regular shelf-stocking visits to shops look like this.
At work on: Periodic vehicle routing
LTL Less-than-truckload
Shipments too small to fill a truck, so several customers share one. The opposite is full truckload (TL or FTL). With LTL you combine several shipments and plan a route with several stops. With TL the main job is to find the right truck for each load.
Sources
- Powell, W. B., Sheffi, Y. (1989). Design and implementation of an interactive optimization system for network design in the motor carrier industry. Operations Research 37(1).
- Crainic, T. G., Laporte, G. (1997). Planning models for freight transportation. European Journal of Operational Research 97(3).
At work on: Trucking and fleet
EV Electric vehicle
For routing, an EV adds a state of charge: the share of battery left. Range drops with load, cold and hills, and charging takes time at specific places. A route that works for a diesel can be impossible for an EV, so the model must track the battery stop by stop.
At work on: Green routing and carbon minimisation
GHG Greenhouse gas (GHG Protocol)
The GHG Protocol is the common standard for counting emissions. Fuel burned in your own vehicles is Scope 1, electricity you buy is Scope 2, transport by carriers is Scope 3. A routing model can minimise emissions directly, so you see what each tonne of CO₂ saved costs.
At work on: Green routing and carbon minimisation
Optimisation methods
The methods behind the plans. You don’t need to run them, but you should know what a vendor means when they name one.
OR Operations research
The field that uses maths to make better decisions about limited resources: routes, schedules, stock, budgets. It grew out of wartime logistics planning. Optimisation is its main tool, and it’s what we do.
Sources
- Churchman, C. W., Ackoff, R. L., Arnoff, E. L. (1957). Introduction to Operations Research. Wiley.
- NEOS Guide: an introduction to optimization models, problem types and solvers (Wisconsin Institute for Discovery).
At work on: Python optimisation development , Excel-based optimisation , Mathematical modelling and OR consulting , Flexible job shop scheduling
Objective function and constraints
Every optimisation model has two parts. The objective is the number you want as low or as high as possible, such as total cost. The constraints are the rules every plan must keep, such as vehicle capacity or rest times. Most arguments about a model are really about what belongs in which part.
Sources
- Williams, H. P. (2013). Model Building in Mathematical Programming (5th ed.). Wiley.
- NEOS Guide: an introduction to optimization models, problem types and solvers (Wisconsin Institute for Discovery).
LP Linear programming
A model where everything grows in step: ten kilometres cost ten times as much as one. There are no volume discounts and no jumps. The decisions can take any value, including fractions. A solver handles an LP fast, even with millions of decisions. The catch is that you can’t send half a truck.
At work on: MILP modelling , Solver performance tuning , Excel-based optimisation , Optimiser not delivering results
MILP Mixed-integer linear programming
An LP where some decisions can’t be split or are a plain yes or no: open this depot, use this truck, run this job first. That’s what most planning in practice needs. A MILP solver gives you a plan and also a bound. The bound is a value that even the best plan can’t beat. So you know how far your plan is from the best possible one, at most. Often shortened to MIP.
At work on: MILP modelling , Constraint programming , Mathematical modelling and OR consulting , Production scheduling
LP relaxation
The same MILP with the whole-number rules switched off, so half a truck is allowed. This simpler calculation is quick, and its result is a bound: no workable plan can be better. Solvers redo it at every step. When the bound is close to the best workable plan, the relaxation is called tight. Then the solver finishes sooner.
At work on: MILP modelling , Optimiser not delivering results
Branch and bound
How MILP solvers search. They split the problem into cases, for example: this depot open, or this depot closed. For each case they first work out how good a plan in it could be at best. To do that, they allow split decisions, such as half a truck. This calculation is called the LP relaxation. If even that best value is worse than the best plan found so far, the whole case is dropped. That way most of the options never get checked one by one.
At work on: Excel-based optimisation , Job shop scheduling , Flow shop scheduling
Optimality gap (MIP gap)
While a MILP solver runs, it shows two numbers: the best plan so far and the bound no plan can beat. The gap is the distance between them. A gap of 2% means the plan is at most 2% worse than the best possible one. You decide when the gap is small enough to stop.
At work on: Solver performance tuning , Optimiser not delivering results
CP Constraint programming
A method built for rule-heavy problems: shift patterns, sequences, “never two night shifts in a row”. The model first strikes out every combination that violates a rule. Only then does it look for the best plan among the rest. CP is strong where keeping every rule matters more than getting the lowest cost.
Sources
- Rossi, F., van Beek, P., Walsh, T. (eds.) (2006). Handbook of Constraint Programming. Elsevier.
- Google OR-Tools: constraint optimization.
At work on: Constraint programming , Mathematical modelling and OR consulting , Workforce shift scheduling , Crew scheduling
CP-SAT solver
A solver from Google for problems with many rules. It’s part of the free toolkit OR-Tools. It combines two methods. The first is constraint programming, which strikes out every combination that violates a rule. The second comes from SAT solvers. Those are programs that solve huge logic puzzles made of yes-or-no questions. It’s free and among the strongest tools for scheduling and rostering.
At work on: Constraint programming , Python optimisation development , Optimisation API development , Optimiser not delivering results
NP-hard
A class of problems for which no fast method is known that always finds the best answer. Add a few stops or jobs and the number of options explodes. Routing and scheduling are NP-hard. That doesn’t mean unsolvable. It means you need good methods and an honest figure for how far the plan is from the best one, at most.
At work on: Job shop scheduling , Flow shop scheduling , Batch scheduling and changeover minimisation , Flexible job shop scheduling
Heuristic and metaheuristic
A heuristic is a rule of thumb that finds a good plan fast, without proof that it’s the best. A metaheuristic keeps improving such plans. It changes a plan a little, checks whether it got better, and repeats that many times. Well-known ones are tabu search, simulated annealing and genetic algorithms. These methods handle huge problems, but they can’t tell you how far their plan is from the best one.
At work on: Custom algorithm development , Large-scale optimisation , Mathematical modelling and OR consulting , Vehicle routing optimisation
LNS Large neighbourhood search
A search method that improves a finished plan piece by piece. It takes one part out of the plan, say ten routes in one area. It plans that part again with an exact method. Exact means the method finds the proven best plan for that part. Then it repeats with another part. It scales to very large problems and keeps much of the quality of exact methods.
At work on: Job shop scheduling
Stochastic optimisation
Planning for several possible developments instead of one forecast. You give the model scenarios, such as low, normal and high demand. The model tests every plan in all scenarios. It picks the plan that costs the least across all scenarios together. Two-stage models decide some things now (how much to order) and some later (how to ship it), once you know more.
At work on: Stochastic optimisation , Safety stock and reorder point , Maintenance scheduling , Perishable inventory optimisation
VSS Value of the stochastic solution
What planning for the range is worth to you. Take the plan built for average demand and replay it across your scenarios. Do the same with the stochastic plan. The difference in cost is the VSS. If it’s small, a plain forecast-based plan is good enough.
At work on: Stochastic optimisation
Pareto front
When two goals work against each other, such as cost in francs and delivery time in hours, there’s no single best plan. The Pareto front is the set of plans where you can’t improve one goal without giving up on the other. We show you these plans as a curve, and you pick the plan that suits you. The model calculates the plans on the curve with the epsilon-constraint method. It sets an upper limit for one goal, say a delivery time of at most 48 hours, and finds the cheapest plan for it. Then it lowers the limit and calculates again.
At work on: Multi-objective optimisation , Multi-objective resource planning , Portfolio optimisation
IIS Irreducible inconsistent subsystem
When a model says “no plan exists”, you want to know which rules clash. An IIS is a smallest set of rules that can’t all be kept at once: drop any one of them and the clash goes away. Solvers such as Gurobi compute it for you, so you don’t hunt by deleting rules one at a time.
Sources
At work on: Optimiser not delivering results
Solvers and tools
The software that does the solving. We pick it per problem, and the code you get works with more than one.
Solver
The engine that takes a model and finds the plan. The model says what you want, the solver works out how. Commercial solvers are often faster on hard problems, free ones are often fast enough. Because we keep model and solver apart, you can switch later.
Gurobi
A commercial MILP solver, among the fastest available. It needs a paid licence for business use, with free licences for academics. Its Python interface is called gurobipy.
At work on: Custom algorithm development , MILP modelling , Python optimisation development , Optimisation API development
CPLEX IBM CPLEX
A long-standing commercial MILP solver from IBM. Many older planning systems run on it. It also has a constraint programming engine.
At work on: Custom algorithm development , MILP modelling , Solver performance tuning , Large-scale optimisation
HiGHS
A free, open-source LP and MILP solver from the University of Edinburgh, under the MIT licence. You can ship it inside your own product without licence fees. For many mid-sized models it’s all you need.
Sources
At work on: Custom algorithm development , MILP modelling , Python optimisation development , Optimisation API development
CBC COIN-OR CBC
An older free MILP solver from the COIN-OR open-source project. It powers OpenSolver in Excel and used to ship with PuLP. It’s slower than HiGHS on most models but still widely used.
At work on: Python optimisation development , Excel-based optimisation
SCIP
A free solver from the Zuse Institute Berlin, strong for MILP and for models with non-linear parts. Researchers use it to build their own methods on top.
At work on: MILP modelling
OR-Tools Google OR-Tools
Google’s free optimisation toolkit. It includes the CP-SAT solver, a routing library and links to MILP solvers. We use it a lot for scheduling and routing.
At work on: Custom algorithm development , Constraint programming , Python optimisation development , Optimisation API development
PuLP, Pyomo and CVXPY
Python libraries for writing models. You describe variables, goal and rules in Python, and the library hands them to a solver. PuLP is the simplest of the three. Pyomo can do more, for example non-linear models. Those are models where cost doesn’t grow in step with quantity. CVXPY is built for one class of problems, called convex. Portfolio optimisation is one example. Recent PuLP versions no longer bundle a solver, so you install one yourself.
At work on: Stochastic optimisation , Python optimisation development
LPL modelling language
A modelling language our founder built at the University of Fribourg. A model in LPL reads close to the maths and runs with several solvers. We use it next to Python where it makes a model clearer.
At work on: Custom algorithm development
MPS file format
An old, plain text format for LP and MILP models that nearly every solver reads. Exporting a model to MPS lets you test it on another solver without rewriting it. In production planning, MPS also stands for master production schedule. That’s the plan that says which products you make, when and how many. So check the context.
At work on: Python optimisation development
GRG Excel Solver, GRG Nonlinear and OpenSolver
Excel’s built-in Solver has three engines. Simplex LP is for linear models. GRG Nonlinear is for models with smooth curves. It can stop at a value that only beats the values right next to it, not all of them. Evolutionary is for everything else. Excel’s Solver limits how many cells it may change. OpenSolver is a free add-in under the GPL licence. It lifts that limit and runs the CBC solver inside the same workbook.
Sources
At work on: Excel-based optimisation , Resource allocation and assignment , Replace your Excel-based planning
VBA Visual Basic for Applications
The macro language inside Excel. Many planning tools started as a VBA macro one person wrote years ago. It works until that person leaves, which is often when we get the call.
At work on: Excel-based optimisation
API Application programming interface
A defined way for two systems to talk. We can deliver your model as an API: your ERP or app sends the data, the model sends back the plan. You don’t have to export spreadsheets by hand.
At work on: Optimisation API development , Real-time dispatch