learn what you must.
solve the rest exactly.
Graybox optimization combines both. kint fits the parts only past data can show, each with its uncertainty and a holdout score. It then solves the decision exactly, in a model your experts can read. A decision with several levels stays one model, with the coupling between levels written down.
fit what history knows. solve the rest exactly.
Most real decisions mix the two. Berth windows and crane counts are rules you can write down. How fast a crane works on a given vessel is something only your logs know.
estimate what only history knows
Each learned part is fitted from the data, scored on data it has not seen, and carries its uncertainty. Gaps are filled and marked as estimates.- Inputrate(v) = fitted rate, lower edge of interval
- Rulemoves(v) ≤ rate(v) × crane hours(v)
- Rulecrane hours(v) ≤ window(v) × cranes(v)
- Rulecranes in use(t) ≤ cranes available
- Goalminimize cranes + carriers in use
solve the decision exactly
The fitted parts enter an exact model as inputs. The solver searches every plan the rules allow and reports bound and gap.berth, then crane, then yard. coupled in one model.
Many decisions are nested. kint solves them level by level and writes the coupling between levels down as constraints. A change on one level then shows up on the others.
plans that hold, trade-offs you can see.
a decision that holds when history cannot pin the future down.
The plan is solved against the range of futures the data allows, so a staffing plan survives a demand spike instead of breaking on it. The fitted intervals from the graybox step feed straight in.
the frontier, then your choice.
Cost against service level: kint returns the set of best compromises. Your team picks the point; the record keeps why.
answers at their own level.
Each model answers at one level: terminal, shift, vessel, crane or move. A question asked at another level gets a named refusal, with the reason, instead of a number that only looks right.
- terminalThroughput per week
- shiftCranes and carriers per shift
- vesselMoves and crane hours per vessel
- craneSequence per crane
- moveSingle container moves
When does crane 3 lift box 41 on vessel B?
Refused: grain mismatch. This model answers per shift and vessel. The question asks per move. A sequencing model at crane level answers it.
one platform, the right family for each part.
kint picks the family with your team, part by part. Every family returns the same answer format and lands in the same record.
| Family | You start from | The result | Example | Page |
|---|---|---|---|---|
| whitebox | Rules you can write down | Optimal for the model as stated once the gap reaches 0.0 %, bound and gap stored | Next week's production mix from live bills of materials in the planning system | whitebox → |
| blackbox | Past decisions and outcomes | Best found on the learned model, with an uncertainty range | Tuning process settings when one simulation run takes hours | blackbox → |
| graybox | Rules plus parts only history knows | Exact solve on fitted parts, each with uncertainty and holdout score | Crane productivity learned, then the shift plan solved exactly | this page |
| robust | A future history cannot pin down | A decision that holds across the range the data allows | A staffing plan that survives a demand spike | above ↑ |
| hierarchical | A nested decision | Each level solved, coupling explicit | Berth, then quay crane, then yard | above ↑ |
| multi-objective | Goals that pull against each other | The trade-off frontier to choose from | Cost against service level | above ↑ |
| simulation | A plan and what could happen | Outcomes per scenario, compared | The shift plan tested against a failed crane | simulation → |
bring a decision with a part nobody can write down.
Bring the rules your planners know and the history that knows the rest. We fit what has to be learned, solve what can be solved exactly, and show you both.
- Families
- whitebox · blackbox · graybox · simulation
- Fitted parts
- Uncertainty and holdout score per part
- Gap
- Bound and gap, stored with the run
- Contact
- hello@kint.pro