AI Glossary / AI and machine learning / Advanced
Stochastic program
Definition
A stochastic program is a statistical model expressed as a procedure that makes random draws. Running the procedure generates a possible outcome, and repeating it explores the range of outcomes implied by the model.
What it means
Some processes contain uncertainty at several connected steps. A shop does not know exactly how many customers will arrive, which will buy or what each will spend. A stochastic program can describe those uncertain quantities and how they relate.
Instead of presenting only a single formula or final average, the program provides a recipe for generating one possible version of events. Repeated runs produce a collection of modeled possibilities.
The randomness is deliberate and defined by probability distributions. It does not mean that the code is carelessly written or that any answer is acceptable. The model should make its assumptions explicit enough to inspect and test.
In this entry, the term follows its statistical-modeling meaning in the supplied source. Stochastic programming can also refer to optimization under uncertainty, which is a related but different usage.
How it works
The modeler defines uncertain variables and their dependencies. A simple procedure might draw daily demand, then draw order sizes conditional on that demand and calculate the resulting workload.
Each execution samples values according to those rules. Many executions provide an empirical picture of the outcomes implied by the model, such as an average, a range or the frequency of exceeding capacity.
Generating samples is different from fitting the model to evidence. Probabilistic-programming systems can add inference procedures that update beliefs about unknown quantities using observations, but the sampling recipe alone does not establish that its assumptions match reality.
Validation includes checking data, dependencies and sensitivity to assumptions. Increasing the number of runs reduces some sampling noise; it does not correct an unrealistic model or include an event the model cannot represent.
Three practical examples
- Staffing exploration: a team simulates uncertain arrival volumes and service times to investigate possible queues.
- Inventory planning: analysts generate demand and delivery-delay scenarios to explore how often a proposed stock policy runs short.
- Project planning: a team models uncertain task durations and dependencies to examine a range of possible completion dates.
Think of it this way
Think of a board-game rulebook that tells players when to roll dice and how each result changes the next move. One playthrough is one possible history; repeated playthroughs reveal patterns allowed by the rules.
The analogy has a limit. In a business model, the dice and rules must be justified by evidence or clearly stated assumptions. Convenient probabilities do not become realistic merely because the simulation runs successfully.
Worked example
Generating an uncertain daily workload
Suppose a teaching model assumes a small service desk receives either 20 or 40 requests, with equal probability. It also assumes every request requires either five or 15 minutes, independently and with equal probability.
One simulated day draws the request count, draws a duration for each request and adds the durations. Repeating that procedure creates different daily workloads.
Under these invented assumptions, expected requests equal 30 and expected minutes per request equal ten, so expected total work is 300 minutes. That average does not describe the whole distribution or prove that a five-hour shift is sufficient.
Real arrivals may cluster, and complicated requests may become more common on busy days. The team would need to examine those dependencies before using the model for staffing. The example demonstrates a sampling procedure, not operational advice.
Illustrative case study
A fictional repair business simulates completion times using independently sampled task durations. The model suggests ample capacity, but real delays often occur together when a shared supplier is late.
The analysts add a common supplier-delay variable that affects several tasks in the same simulated run. They compare the revised outputs with historical patterns and discuss which assumptions remain uncertain.
The lesson is that relationships between uncertain events can matter more than the number of simulations. A thousand runs of an independence assumption cannot reproduce a shared disruption that the model has excluded.
Common misconceptions and mistakes
- Stochastic means the software is unreliable. The randomness is part of the specified model and can be handled systematically.
- More simulations make the assumptions true. More runs refine the model's numerical picture; they do not validate its connection to reality.
- A simulated range includes every possible outcome. Unmodeled events and unsuitable distributions can leave important possibilities out.
Business use
Stochastic programs help explore uncertainty when a single average hides meaningful variation. They can support discussions about capacity, resilience and planning assumptions.
Ask what is random, which quantities depend on one another and how distributions were chosen. Review sensitivity to plausible changes before treating a result as a sound basis for action. Keep simulated outputs clearly distinguished from observed performance.
People also ask
Can a stochastic program be reproduced?
A fixed random seed and controlled environment often make a run reproducible, although exact behavior can depend on the software and hardware.
Is every simulation a trained AI model?
No. Many are hand-specified statistical procedures with no learned component.
Does it automatically find the best decision?
No. Simulating outcomes and optimizing a decision are separate tasks, although they can be combined.
Check your understanding
A simulation omits a supplier disruption affecting several tasks. Will running it a million more times solve that omission?
Answer: No. The model must represent the shared disruption or a justified alternative. Extra runs only explore the possibilities already included in its rules.
Related terms
Learn next
Read prior belief and Monte Carlo integration. Write down one uncertain process and distinguish observed data, assumed probabilities and dependencies before modeling it.
More in AI and machine learning
- Statistical relational learning - Statistical relational learning is a family of methods that learn and reason under uncertainty while representing relationships among entities.
- Step size - In optimization, step size controls how far an update moves along a chosen search direction.
- STIT logic - STIT logic is a family of formal logics for reasoning about agency: an agent “sees to it that” some statement is true.
- Stochastic variance reduction - Stochastic variance reduction is a collection of optimization techniques that reduce randomness in estimated gradients while retaining much of the...
- Streaming Feature Pipeline - A streaming feature pipeline continuously processes incoming events into features that machine-learning systems can use.
- Stride - Stride is the distance a sliding operation moves between successive positions in an input.
Sources
Thomas Minka, A Statistical Learning/Pattern Recognition Glossary, stochastic program: https://alumni.media.mit.edu/~tpminka/statlearn/glossary/glossary.html. Pyro, An Introduction to Models in Pyro (archived introductory tutorial): https://pyro.ai/examples/intro_part_i.html. Checked 12 October 2026. Scenarios are original.
Previous: Stochastic process Next: Stochastic semantic analysis