- Home
- AI development
- Reinforcement learning
Reinforcement learning
Reinforcement learning, for the few decisions where it earns its cost
Reinforcement learning trains a policy by trial and error in a simulator, scoring each outcome by the measure you care about. For a small or mid-size business it is rarely the first tool, because most repeated decisions are solved faster with an optimization model or a plain rule, and we say so before we build. It fits decisions that repeat often, show their result quickly and can be simulated cheaply, such as dispatch order, scheduling and reorder points.
- Prices published on the site
- First working version in two weeks
- Support month to month, cancel any time
A policy that learns by trying, in a simulator before it touches live orders
In reinforcement learning, a program picks an action, sees a score and adjusts so the next pick scores better. Repeat that millions of times against a simulator of your operation and you get a policy, a function that says which job goes to which driver next or when to reorder a part. The simulator is built from your own history: past orders, arrival times, delays and stock levels.
The method needs three things at once. The decision repeats often, you learn the outcome within days rather than quarters, and your history is good enough to simulate. Without all three, a simpler method wins. The same family of methods, reinforcement learning from human feedback, is how language models are tuned to the answers people prefer, and that work is on LLM fine-tuning.
Simulate, act, score and compare against your current rule
Simulate
A simulator of the decision, built from your history. Before anything else it has to reproduce last year's numbers.
Act
The policy tries actions in the simulator, far faster than real time, so it meets the bad days at no cost to you.
Score
Each outcome is scored by the measure you care about, such as miles driven, late deliveries, overtime hours or stockouts.
Compare
The policy is compared with your current rule, or a simple heuristic, on held-out history it never trained on. It goes live only if it wins on a final set of history kept apart until the end.
Before any build, the written review checks whether a simple rule or a solver is enough. If a project goes ahead, the stopping point and what you receive are written into the scope.
Four decisions, and what we try first for each
Each of these can suit reinforcement learning. Each also has a cheaper method that we test first, because it often wins.
Dispatch order
Which job goes to which driver or crew next, when jobs arrive through the day and travel time matters. We try a routing solver such as Google OR-Tools first.
Scheduling
Shifts and appointments where cancellations and no-shows change the picture hour by hour. We try a constraint solver over your shift rules first.
Reorder points
When and how much to reorder for parts or supplies with uneven demand and late suppliers. We try a reorder formula from last year's usage first.
Queue priority
Which request in a queue to take next when waiting time and effort both count. We try oldest first, then highest value first, and measure both.
When reinforcement learning pays off, and when something simpler does
Worth building when
- The same decision is made hundreds of times a month
- The result shows within days, so the score arrives quickly
- You have a year or more of records of what was decided and what happened
- A rule or a solver has been tried and leaves money on the table
Use something simpler when
- The decision comes up a few times a month, and a person with a spreadsheet does as well
- The problem is one schedule or one route, which an optimization model solves directly
- You have no records of past decisions to build a simulator from
- Every answer must be explained rule by rule, which is the job of expert systems
The simulator is most of the work
An optimization model or a scheduling solver for one decision.
Several systems made to work as one: integrations, roles, reports.
A simulator, a trained policy and its comparison with your current rule.
Real reinforcement learning, with a simulator built from your history and a held-out comparison, usually lands in a full build, from $15,000. An optimization model or a scheduling solver for the same decision is often one process. The audit tells you which one your problem needs, and sometimes the answer is a rule that costs less than either.
Care $300 to $500 a month. Active $750 to $1,500 a month. AI model usage, GPU time and hardware are paid to the provider, on accounts in your name.
Questions about reinforcement learning
Is reinforcement learning right for a small business?
Rarely. It needs a decision that repeats often, a quick result and history to simulate from. Dispatch, scheduling and reorder points can qualify. In most cases an optimization model or a written rule gets most of the gain at a lower price, and its answers are easier to read.
What is the simulator, and who builds it?
A program that imitates your operation closely enough to test decisions on. We build it from your history, such as past orders, delays and stock counts, and check that it reproduces last year's results before the policy trains in it. You can read how it works.
How much history do we need?
At least a year of records that show what was decided and what happened next, so the simulator covers every season. With less, or with gaps, we say so in the audit and suggest a simpler method.
What does it cost to run once it is built?
Training uses cloud compute or GPU time, billed by the provider to an account in your name. A finished policy for a decision like dispatch or reordering usually runs on an ordinary server. We estimate both numbers in the audit.
Is this the reinforcement learning used to tune chatbots?
It is the same family of methods with a different job. Reinforcement learning from human feedback tunes a language model toward the answers people rate higher, and we cover it on LLM fine-tuning. Here a policy picks actions in your operation and is scored in a simulator.
Related services
- Expert systems Your rules, applied the same way every time
- LLM fine-tuning Train a model on your own examples
- MLOps Run, monitor and update models in production
- LLM evaluation Test which model does your job best
- AI readiness assessment A written review of where AI fits
- AI development Every AI service we build, on one page
Tell us what's slowing you down.
You get a free written review by email within two business days.