Optimisation software: models that fit plants, solvers that hold

Production models with capacities and costs must really optimise

What happens. A vendor of decision software builds a model of the customer’s production with machines, capacities, set-up times and costs, and the model should find the best plan. In practice the model is a simplification, the data are late, and the plan that comes out is one the planner corrects by hand.

What it costs. Software projects that take long to bring benefit, customers who go back to spreadsheets and consultants who spend their time on data and not on the model.

Why the usual fix fails. A more detailed model is slower and fragile. A simpler one loses the very effects that matter. Heuristics tuned for one customer need to be rebuilt for the next. Hiring a mathematician for each customer is not scalable.

How a fixed-price remote diagnosis checks it. From a typical customer model, its data and one planned example I check which simplifications cost most, whether the solution method suits the structure, and where the solving time goes. The fixed-price remote diagnosis answers whether a better formulation or method brings the plan close enough to what the planner wants.

Airline planning needs simulation and fast optimisation heuristics

What happens. Airlines plan flights, crews and aircraft in a network where one delay spreads. Planners want to simulate the effect of changes and find a better plan in minutes, not in a night’s run. The tools that exist are slow or give a plan that does not survive the first disruption.

What it costs. Delays that cost compensation and crew hours, aircraft on the ground and planners who work late to repair the plan.

Why the usual fix fails. Adding buffer to the schedule protects against delay and wastes aircraft time. An exact method takes too long for the size of the network. A rule of thumb does not see the second effect of a change.

How a fixed-price remote diagnosis checks it. From a sample of a plan and its disruptions I check which simulation detail matters and which heuristic gives a plan good enough within your time limit. The fixed-price remote diagnosis answers what the computation time and the quality of the plan would be for your network.

Background: Optimal paths and setpoint plans

Optimisation-based driving must scale with a safety argument

What happens. A vehicle that is driven by an optimiser has to deliver a safe move every fraction of a second. For a pilot with a few vehicles the solver is tuned by hand. For a fleet it must run on the controller of the vehicle, with a safety path that an assessor can accept. Customers of the software ask for help with the solver itself.

What it costs. Delays when moving from pilot to series, support effort for each customer and a risk that the solver fails in a rare case.

Why the usual fix fails. A solver tuned for the test track fails in the rare case. A second, simple fallback controller secures the vehicle but does not remove the question of when the optimiser is trusted. A general solver library is too big for the vehicle’s controller.

How a fixed-price remote diagnosis checks it. From the formulation of the driving problem, solver settings and failure cases I check where solving time and fail cases come from and what a lean solver with a documented fallback needs. The fixed-price remote diagnosis answers whether the solver can be made robust enough for a series release.

Background: Predictive control on embedded hardware