Hard and soft scheduling constraints belong in different parts of the solver
Hard scheduling constraints make a schedule invalid, so non-overlap, table capacity and declared availability belong in the feasibility model. Soft constraints make a schedule annoying, so language preference and idle time belong in a weighted objective. Reporting violations of the two together destroys the information you needed.
The schedule comes out with a total penalty of 412 and the team spends an afternoon arguing about whether 412 is good. Nobody can say, because 412 is the sum of a language mismatch weighted at 3, a long gap weighted at 5, and one participant booked into two meetings at 14:20 weighted at 200 because somebody wanted that to be strongly discouraged.
Hard and soft scheduling constraints got folded into one objective function, and the result is a schedule that is arithmetically excellent and operationally impossible. The double booking is still there. It was priced instead of prevented, and the price was paid by an exhibitor standing at table 14.
The distinction the literature already drew
Constraint satisfaction has a precise version of this, and it is worth borrowing the vocabulary because it settles arguments quickly.
A constraint satisfaction problem asks whether an assignment exists that satisfies every constraint. Freuder and Wallace, in Artificial Intelligence in 1992, took up what to do when no such assignment exists, which is the normal state of a meeting programme. Their partial constraint satisfaction framework weakens the problem deliberately: you decide in advance which constraints may be relaxed and by how much, then look for the best solution to the weakened problem. The decision about what may be relaxed comes first, before any search runs.
Schiex, Fargier and Verfaillie took that further at IJCAI in 1995 with valued constraint satisfaction, attaching a cost from an ordered structure to each violated constraint. Their framework makes room for costs that are genuinely infinite, meaning a constraint whose violation can never be traded against any amount of gain elsewhere. That is precisely what a hard constraint is, and expressing it as a very large finite penalty is a different thing that behaves differently under optimisation.
A penalty of 200 for a double booking is not infinity. If the solver can gain 250 points of match score by accepting one, it will, and it will be right to under the objective you wrote. The failure was in the objective.
What belongs on each list?
The split is not a matter of taste. Ask one question of each rule: if this is broken, does the schedule still run?
Hard. Participant non-overlap, because a person cannot be in two places. Table capacity per slot, because tables are physical objects. Declared availability, because a buyer in a taxi cannot attend. Forbidden slots, such as a religious observance or a keynote a buyer is speaking at. Fixed meetings, where a contract says this meeting happens at 11:00 at table 1. Nguyen, Kieu and To, in their 2026 arXiv work on compact encodings for business to business meeting scheduling, keep session, fixed-slot, forbidden-slot and precedence restrictions inside the slot domain of each meeting, so an infeasible placement cannot be represented at all.
Soft. Idle gaps between meetings. Language preference between the two parties. Meeting order, where a buyer would rather see the three flooring suppliers consecutively. Proximity, where a table near the buyer's last meeting beats one across the hall. Balance across exhibitors. Every one of these can be broken and the show still happens.
The two lists have different owners in practice. The hard list is owned by operations, because operations is who discovers a violation. The soft list is owned by whoever is responsible for the participant experience, and their weights are a policy choice they should be asked to defend in writing.
One test settles most disputed cases. Imagine the schedule has been published and the rule turns out to have been broken for one participant. If the correct response is to reissue the schedule, the rule was hard. If the correct response is to apologise and carry on, it was soft. A show director who says both, that they would reissue and also carry on depending on the participant, has told you the rule is really two rules with different populations, and it should be modelled that way.
How do you handle a constraint that is nearly hard?
Some rules genuinely sit in between, and pretending otherwise causes its own damage.
The common one is a sponsor's guaranteed meeting count. Contractually it is hard: the sponsor paid for 12 meetings. Physically it is soft, because if only 9 buyers in the room want a meeting with them, no solver can conjure the other 3.
Handle it in two stages. Encode the sponsor's meetings as hard up to the number that is actually achievable, computed before the solve from the count of accepted proposals, and put the remainder into the objective at a high weight. Then report the shortfall as its own line in the output. The commercial team needs to know that the guarantee failed at 9 of 12 and why, and they need that as a fact rather than as a component of a penalty total.
The second in-between case is availability. A buyer who declared availability from 11:00 could probably make 10:40 at a push. Treating declared availability as hard is still the right call, because the alternative is a solver quietly deciding which buyers get inconvenienced. If you want the flexibility, widen the declared window explicitly and say you have done it.
The general rule I would apply is that a constraint may be softened by a person, in the data, before the solve. It may not be softened by the solver during it. That keeps every relaxation visible in an input file somebody can read.
Report the two separately, always
Here is the reporting format that ends the arguments.
Feasibility comes first as a pass or fail with a count of violations by type. A schedule with any hard violation is not a candidate. If your solver is returning schedules with hard violations, it is not modelling them as hard, whatever the config file says.
Quality comes second, decomposed by soft constraint type, in the units of the constraint rather than in penalty points. Take a run over 200 buyers and 1,860 meetings. Report it like this: mean internal idle slots 2.4 and worst case 20; language mismatches 148 meetings, which is 8.0 per cent; consecutive-run preference honoured for 61 of 200 buyers, which is 30.5 per cent; exhibitor meeting counts ranging from 0 to 19 with a median of 7.
Nobody needs a penalty total to read that. A show director can look at the worst case of 20 idle slots and the exhibitor range topping out at 19 and know exactly which two problems to fix, and the weights that produced this schedule are now visible as consequences instead of as inputs.
The weights still exist inside the optimiser. Keeping them out of the report means next edition's numbers stay comparable when somebody changes a weight, which they will.
The failure mode this prevents
Folding hard rules into the objective produces a specific and recognisable disaster, and it usually surfaces about 48 hours before doors.
Somebody raises a weight to make the solver care more about a soft preference. Say idle time goes from 5 to 25 because buyers complained last year. The solver, now heavily penalised for gaps, starts finding it profitable to place a meeting into a slot where the buyer already has one, because 200 points for the overlap is less than the 25 points per idle slot it saves across a compacted day of ten participants.
Nobody sees this in testing, because the test schedules were smaller and the trade never became profitable at that size. It appears on the full run, on the last rebuild before publication, and it looks like a bug in the solver. It is not. It is the model doing what it was told, and it is exactly the case Schiex and colleagues had in mind when they insisted that some costs cannot be aggregated with others.
Once non-overlap sits in the feasibility model, this cannot happen at any weight. That is the entire argument for the separation, and it is why enforcing non-overlap at write time matters just as much for the bookings that arrive outside the solve.
Where this stops
A clean split does not tell you the weights. The soft constraints still trade against each other, and there is no principled way to know whether one idle slot is worth more or less than one language mismatch. That number comes from your programme, your participants and your commercial position, and anybody who hands you a default is guessing.
The second limit is that hard constraints can make a problem infeasible, and then you get nothing. On a tight floor, a solver told that capacity, availability and non-overlap are all inviolable may return no schedule whatsoever, and the honest response is to find which constraint to relax in the data rather than to soften them all in the model. Drop the meetings that cannot be placed, and put them on a standby list where somebody can see them.
The third is that a constraint you cannot measure cannot be either. Language preference needs a language field on both sides of the pair, and if 40 per cent of your exhibitor records have none, the soft constraint is operating on the 60 per cent that do and quietly favouring the exhibitors with complete profiles. Report coverage next to every soft constraint, because a preference computed on partial data is a preference for good data entry.
Start by listing every rule your scheduler currently applies, in one column, and marking each one as hard or soft against the single test of whether the show can still run. Then check whether your first-fit placement pass and your solver agree on that marking, because they usually do not, and reconcile them before the next run. Any rule marked hard that lives in the objective, or soft that lives in the feasibility model, is a defect worth fixing before you tune a single weight in your matchmaking configuration, and the same list tells you which numbers belong in the idle time report.
Questions people ask about hard and soft scheduling constraints
- What is the difference between a hard and a soft scheduling constraint?
- A hard constraint defines whether a schedule is valid at all. Booking one person into two meetings at once or using more tables than the hall holds produces a schedule that cannot be run. A soft constraint expresses a preference: shorter gaps, a shared language, a morning start. Breaking one costs quality without breaking the schedule.
- Why not put every constraint into one weighted penalty function?
- Because a single number cannot tell you which rule was broken. A schedule scoring 412 penalty points might have one double booking or forty long gaps, and those need different responses from different people. Separating feasibility from objective means an infeasible schedule is rejected rather than priced.
- Which meeting scheduling constraints should be hard?
- Participant non-overlap, table or location capacity per slot, declared availability, forbidden slots, and any meeting fixed to a particular time by contract. Everything about the shape or pleasantness of a day, including idle gaps, meeting order and language matching, belongs in the objective as a weighted preference.
Related reading
- Idle time in meeting schedules is the complaint you will hear
- Double booking prevention in a meeting programme with three sources of truth
- Why a greedy meeting scheduler is often the right first build