AI Glossary / Data science and analytics / Intermediate

Data farming

Definition

Data farming is the purposeful generation of data from simulation models through designed experiments. Analysts vary model inputs, run selected combinations, and study the resulting outputs to learn how the simulated system behaves. The aim is informative exploration, not simply producing the largest possible dataset. Findings remain conditional on the model and the assumptions represented in it.

What it means

Data mining often starts with data that already exist. Data farming emphasizes deciding which simulation runs to perform so that the generated data address a question. A team can explore combinations that are expensive, rare, or impractical to observe directly. However, simulated observations are not automatically evidence that the real world will behave in the same way. The model's suitability and the experiment's coverage are part of the evidence.

How it works

Define the decision question, model outputs, input factors, and ranges of interest. Choose an experimental design that provides useful coverage within available resources. Run the simulation at the selected design points, using replications where needed to study stochastic variability. Preserve model versions, configurations, and randomization information. Analyze patterns, interactions, and uncertainty, then refine the model or experiment when findings reveal gaps. Efficient experimental designs can explore a space more intelligently than changing one factor at a time or exhaustively testing every possible combination.

Three practical examples

• Service planning: a team simulates different staffing levels, demand patterns, and service-time assumptions to examine waiting-time behavior.
• Logistics design: analysts explore warehouse layouts and routing policies in a validated simulation before considering a real-world trial.
• AI development: a team studies an agent in simulated environments with systematically varied conditions, keeping generated experience distinct from real deployment evidence.

Think of it this way

Think of planting experimental plots with deliberately chosen combinations of soil treatment and watering. The design determines what can be learned, rather than merely how many plants grow. The metaphor is the source of the term's appeal. In data farming, the plots are simulation settings, and the results depend on programmed assumptions rather than physical field observations.

Worked example

Suppose a fictional service simulation has two staffing settings, three arrival-rate settings, and four service-time settings. Testing every combination gives 2 × 3 × 4 = 24 design points. If the stochastic model is run 10 times at each point, the experiment produces 240 runs.
Twenty-four settings and 240 runs are different counts: replication measures variability at the same setting, while additional design points explore other settings. If one staffing level performs well only when arrivals are low, averaging all results into a single staffing score could hide that interaction. The analyst should examine performance across relevant conditions and report uncertainty. More replications reduce some simulation sampling uncertainty, but they do not repair a model that omits an important cause of real-world delays.

Illustrative case study

In a fictional fulfillment company, analysts build a simulation of order processing and use data farming to compare staffing policies across demand scenarios. A promising policy performs well in ordinary conditions but poorly during a simulated surge. The team examines the mechanism and discovers a queue-handling assumption that needs operational review. It updates the model only after checking the real process, then repeats the relevant experiments with clear version records. The example shows an iterative learning process, not a claim that simulated optimization alone validates a deployment decision.

Common misconceptions and mistakes

Data farming is not a synonym for collecting agricultural data or indiscriminately harvesting personal information. Running a model thousands of times does not establish that the model is correct. A simulated causal relationship is a relationship within the model and should not automatically be generalized to the real system. Repeated identical deterministic runs add no stochastic information. Finally, selecting only the best-performing simulation conditions can produce misleading confidence if the intended operation includes a broader range of circumstances.

Business use

Businesses can use data farming to explore alternatives and identify conditions requiring more evidence. Define outputs that reflect the actual decision, including undesirable outcomes and resource limits. Keep experimental design separate from claims about real-world effectiveness. In AI projects, generated simulation data can support development and testing, but teams should assess the gap between the simulated environment and intended use. Document which factors were varied, which were held fixed, and which were omitted. Those boundaries determine what questions the experiment can reasonably answer.

People also ask

Is data farming the same as synthetic-data generation?

It overlaps, but data farming specifically emphasizes designed simulation experiments and analysis of their outputs; synthetic data can be generated for many other purposes.

Must it test every input combination?

No. Suitable experimental designs can select informative combinations efficiently.

Do more runs always improve validity?

No. They may improve numerical precision under the model while leaving structural errors and unrealistic assumptions unchanged.

Check your understanding

A team runs 20 replications at each of 12 simulation settings. Has it explored 240 distinct settings?

Answer: No. It has 12 design points and 240 runs. The replications help study variability at those settings, while the experimental design determines coverage of the input space.

Related terms

• Design of experiments
• Synthetic data
• Sensitivity analysis
• Replication (statistics)
• Model validation
• Data mining
• simulation
• replication
• experimental factor

Learn next

design a small simulation experiment with two factors and a clearly defined output. Separate the number of settings from the number of replications, then inspect whether the effect of one factor changes with the other. Write down one real-world mechanism absent from the simulation before interpreting its results.

More in Data science and analytics

• Cross-industry standard process for data mining - The Cross-Industry Standard Process for Data Mining, usually abbreviated CRISP-DM, is a process model for organizing data-mining projects.
• Dark data - Dark data is information an organization has collected or retained but does not meaningfully use for analytics or decision-making.
• Data cleaning - Data cleaning is the process of identifying and addressing errors, inconsistencies, and other quality problems in data so that the data are more suitable...
• Data physicalization - Data physicalization represents data through the geometry or material properties of physical artifacts.
• Data stream mining - Data stream mining discovers patterns or updates analytical models from data arriving over time, often under limits on memory and processing time.
• Data structure - A data structure is a way of organizing and representing data so that particular operations can be performed on it.

Sources

Sources: Susan Sanchez's original article describes purposeful simulation-data generation: https://doi.org/10.1145/3425398 . NPS's SEED Center lists its primary data-farming and simulation-experiment materials: https://nps.edu/web/seed/papers-presentations . Relevant author and institutional excerpts were consulted, not a complete review of the cited methods. The service simulation arithmetic and fulfillment-company scenario are original civilian teaching examples; no application-specific findings from the research are imported.

Previous: Data fabric Next: Data-flow diagram