Thompson Sampling
Definition
Thompson Sampling randomly selects a hypothesis from the learned distribution for a decision.
A Bayesian strategy for sequential decisions that draws a model hypothesis from the posterior and then acts accordingly.
Simply explained
The method draws a possible explanation fitting the current knowledge and decides based on it. With new results, the basis for later decisions changes.
Why it matters in practice
The term helps to select, develop, evaluate AI systems in a well-founded manner and realistically assess their technical risks.
Example from business
A retailer adaptively distributes limited test traffic across offer variants while maintaining fixed safety limits.
Limits and pitfalls
- Incorrect priors, delayed feedback, and non-stationary effects can bias decisions; exploration remains risky.
Origin & history
William R. Thompson described the method in 1933 for adaptive treatment decisions.
Use in mid-sized companies
- Possible application: A retailer adaptively distributes limited test traffic across offer variants while maintaining fixed safety limits.
Frequently asked questions
What does Thompson Sampling mean?
A Bayesian strategy for sequential decisions that draws a model hypothesis from the posterior and then acts accordingly. What should you watch out for? Incorrect priors, delayed feedback, and non-stationary effects can bias decisions; exploration remains risky.
Sources
- On the Likelihood that One Unknown Probability Exceeds Another · Thompson, On the Likelihood that One Unknown Probability Exceeds Another
Source checked on 10.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: