Description
An adaptive experiment that allocates interventions in proportion to the probability each is best, balancing exploration of new options with exploitation of the best-known one.
Thompson Sampling maintains a posterior distribution over each arm's reward, samples from each posterior, and chooses the arm with the highest sample, so traffic shifts toward better arms while uncertain arms still get explored. In Vivek Gopinath's drift playbook it selects adaptive UX interventions per user (personalised re-onboarding, dynamic UI scaling, context-aware nudging), with the objective of reducing DriftScore, increasing entropy and diversity, and improving CTR. In the e-commerce case it chooses between 'explore collections' and 'deal of the day' per user. Bandit results are then confirmed with A/B tests on DriftScore reduction and entropy lift and with longitudinal interviews on regained delight.
Attributes
- Discipline
- Experimentation & Causal Inference
- Loss type
- Siloed
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Define the candidate interventions (arms) and the reward (e.g. DriftScore reduction, entropy lift, CTR).
- Set priors for each arm's reward (e.g. Beta priors for binary outcomes).
- For each eligible user, sample from each arm's posterior and serve the arm with the highest sample.
- Observe the reward and update the chosen arm's posterior.
- Monitor allocation and guard against spamming users with nudges.
- Confirm the winning policy with an A/B test and qualitative follow-up.
Sources
- Vivek Gopinath (2025), Detecting Behavioral Drift: A Mixed Methods Playbook and E-commerce Case Study
Also known as
- Bayesian bandits
- Multi-armed bandit optimisation
Categories
- Analytics & Behaviour
- Experimentation & Causal Inference