Service
Decision systems: reinforcement learning, simulation and honest evaluation
Simulation, reinforcement learning and honest evaluation for decisions you make again and again, where each wrong call has a cost. A new policy has to beat the rule you use today, on data it has never seen, before it goes anywhere near production.
In three steps
How a decision system is built
Simulate
A simulator of the decision, built from your history of past decisions and outcomes, with every assumption written down.
Evaluate
Today's rule and each new policy run in the same simulator, on held-out periods, with costs, limits and the worst cases reported.
Decide
A policy goes forward only if it beats today's rule. It starts in a trial you control, with limits it cannot cross.
In plain words
What reinforcement learning is
Reinforcement learning is a way to learn a policy, a rule for choosing an action, by trying actions in an environment and scoring what happens. The policy improves by doing more of what scored well over time, not by copying labelled answers.
Trying actions on a live business is expensive, so the trying happens in a simulator built from your history. That is why the simulator, and an honest test of it, come first. Often a simpler method wins: a tuned rule or a standard optimisation solver. When it does, we recommend it.
01 Who it is for
Teams that make the same kind of decision many times: how much stock to reorder, which job to schedule next, how to set a price, where to send a vehicle or a request. Each choice has a cost, and the result shows up later.
02 Signs you need it
- The decision follows a rule of thumb that someone tunes by hand.
- A forecast got better, but the decisions built on it did not.
- There is no safe way to try a new policy before it goes live.
- A policy that looked good in testing did worse than the rule it replaced once it was used.
- Nobody can say what today's rule costs compared with a better one.
03 What you get
- A simulator
- Built from your data, with its assumptions and limits written down.
- A baseline
- Your current rule, measured in the same simulator, so every comparison is fair.
- Candidate policies
- Reinforcement learning where it helps, and simpler methods where they are enough.
- An honest evaluation
- Held-out periods, real costs and agreed limits, with the unflattering numbers shown next to the good ones.
- Guard rails
- Limits that no policy may cross, and a switch back to the current rule.
- Code and a report
- In your repository, so your team can rerun every number.
04 What we need from you
- A record of past decisions and what happened after them.
- The costs and limits that matter: money, capacity, time, risk.
- One person who owns the decision and can say what “better” means.
- A way to try a policy safely, such as a shadow run next to the current rule.
05 What a first pilot looks like
One decision, simulated from your data, with today's rule and one alternative measured on periods the work never trained on. The pilot ends with the simulator, the comparison, and a plain answer: is a new policy worth a trial, or not?
The scope, the price and the length are agreed in writing before it starts. How a pilot works.
06 When we would say no
- A fixed rule or a standard optimisation solver already does the job. Reinforcement learning is often not needed, and we will say so.
- There is no record of past decisions and outcomes, and no safe way to collect one.
- The cost of one wrong action cannot be limited.
Next
A decision that runs on a rule of thumb?
Tell us what is decided, how often, and what a wrong call costs. An engineer replies.