Is ABC the same as choosing the closest simulation?

No. Bayesian ABC represents parameter uncertainty through a distribution, although optimization and ABC can share simulation components.

Does reducing the tolerance always solve the problem?

No. It can reduce one approximation while leaving inadequate summaries or an incorrect simulator unchanged.

Must all ABC algorithms use simple accept-or-reject sampling?

No. There are sequential, weighted and other formulations; their implementation and target distributions need to be understood.

AI Glossary / Statistics and probability / Intermediate

Approximate Bayesian computation

Definition

Approximate Bayesian computation (ABC) is a family of methods for learning about model parameters by comparing simulated data with observations when evaluating the model's likelihood is difficult. Parameters that generate sufficiently similar results receive support. The resulting distribution approximates a Bayesian posterior under a specified model, prior, comparison rule and tolerance.

What it means

Some models are easy to run forward but hard to evaluate backward. A simulator may generate customer arrivals, biological populations or interacting agents without providing a convenient formula for the probability of the observed dataset. ABC uses those simulations as evidence. It does not eliminate assumptions: the simulator and prior still determine which explanations can be considered. Similarity also needs a defensible definition, especially when a complex dataset is compressed into a few summaries.

How it works

In basic rejection ABC, draw a parameter value from the prior and simulate a dataset using that value. Compute the same summary statistics for simulated and observed data, calculate their distance, and retain the parameter if the distance is no greater than a chosen tolerance. Repeat many times. Accepted values represent the corresponding approximate posterior. More advanced algorithms use sequential populations, importance weights or Markov chains to improve efficiency. A smaller tolerance generally demands more computation. Even a zero tolerance does not recover all information if the chosen summaries discard relevant evidence.

Three practical examples

Population research: compare genetic summaries generated under different demographic histories with observed summaries to investigate plausible parameters.
Operations modeling: calibrate an agent simulation against several observed queue and waiting-time features when a tractable joint likelihood is unavailable.
Model development: generate synthetic datasets with known parameters and check whether an ABC procedure recovers them with appropriately calibrated uncertainty before analyzing operational data.

Think of it this way

Think of adjusting a miniature traffic simulator until its behavior resembles measured traffic. Many settings can look similar, and matching only average speed may conceal very different congestion patterns. ABC keeps a distribution of compatible settings rather than declaring that the best-looking simulation proves one unique explanation.

Worked example

Consider an intentionally simple model in which a success probability p is either 0.2 or 0.8, with equal prior probabilities. Observe two independent trials, both successful, and summarize the data by the number of successes. For exact matching, a simulated dataset must contain two successes. Acceptance probabilities are 0.2 squared = 0.04 and 0.8 squared = 0.64. Among accepted draws, posterior support for p = 0.8 is therefore 0.64/(0.04 + 0.64) = 16/17, about 94.1%. Now allow the simulated count to differ by at most one. Counts of one or two qualify. Acceptance probabilities become 1 - 0.8 squared = 0.36 and 1 - 0.2 squared = 0.96. Support for p = 0.8 falls to 0.96/1.32 = 8/11, about 72.7%. These are limiting proportions, not guaranteed finite simulation counts. The example has an available exact likelihood deliberately, so the effect of tolerance can be checked directly.

Illustrative case study

Fictional case: Cedar Queue Analytics calibrates a simulator for a service desk. Matching average waiting time alone produces parameter settings with unrealistic extreme delays. The team adds a waiting-time percentile and a measure of queue persistence, scales the summaries appropriately, and examines several tolerances. It also checks held-out operating periods and whether the simulator can reproduce features omitted from calibration. Managers receive uncertainty ranges and sensitivity results, not a claim that one inferred arrival mechanism is proven. The case illustrates a validation workflow and reports no measured commercial outcome.

Common misconceptions and mistakes

“Likelihood-free” means the method can avoid explicitly evaluating the likelihood; the simulator still implies a data-generating probability model. A high acceptance rate is not automatically good: a loose tolerance can accept uninformative matches. More simulations reduce Monte Carlo error but do not automatically remove summary-information loss, tolerance error or model misspecification. Distances can be dominated by summaries with larger units unless scaling is addressed. Comparing models using inadequate summaries can be particularly misleading. An approximate posterior is conditional evidence, not a causal explanation established by resemblance.

Business use

For business use, ABC can connect sophisticated simulations with observed evidence when simpler estimation methods are impractical. Specify which decisions the parameter estimates will support, what observations the simulator should explain, and how computation costs will be managed. Retain simulator versions, priors, summary definitions, tolerances and random seeds. Compare with a tractable baseline wherever possible. Before using results for investment or staffing decisions, test sensitivity to modeling choices and assess predictive performance on situations that were not used for calibration.

Check your understanding

Q: In the worked example, why does the wider tolerance weaken evidence for p = 0.8?

Answer: A: It admits one-success datasets, which are less distinctive than two successes and give the lower probability model relatively more support. The change is caused by the acceptance rule, not by collecting new observed trials.

Related terms

Bayesian inference
Prior distribution
Likelihood
Rejection sampling
posterior distribution
simulation-based inference
summary statistic
Monte Carlo error
posterior predictive check and model identifiability

Learn next

implement the two-value example and compare empirical acceptance proportions with the exact fractions. Then vary the tolerance and sample size. For a richer simulator, test recovery using synthetic observations and inspect whether summaries distinguish parameter combinations before interpreting results from real data.

Sample size

Sources

Sources: Beaumont, Zhang and Balding, “Approximate Bayesian Computation in Population Genetics,” foundational summary-based methodology: https://pmc.ncbi.nlm.nih.gov/articles/PMC1462356/ ; Aeschbacher, Beaumont and Futschik, “A novel approach for choosing summary statistics in approximate Bayesian computation”: https://pubmed.ncbi.nlm.nih.gov/22960215/ ; Wilkinson, “Bayesian Computation and Model Selection Without Likelihoods,” methodological treatment: https://pmc.ncbi.nlm.nih.gov/articles/PMC2815920/ .

Previous: Approximant sound Next: Approximate entropy