Bandits Lab

Watch strategies learn which slot machine pays best, and what each loses while learning. Or plan an A/B test: see whether a fixed 50/50 split or shifting traffic to the winner costs fewer sales.

Race the strategies

Lower is better. The dashed line is the best any strategy can do in the long run.

Machines

Payout rate of each arm 2–10 numbers between 0 and 1, comma-separated.

Rewards

Rounds Runs averaged

Strategies

Strategy settings

ε (fixed) c (decaying ε) m (ETC) c (KL-UCB) Beta prior a₀ Beta prior b₀ Seed

Same algorithms as the Python library, run in your browser. The random numbers differ, so results match the Python on average, not run for run.

Regret is how many wins you gave up compared with always playing the best arm (± is the standard error over runs). The slope is fitted over the last 90% of rounds; an optimal strategy's slope approaches C as the rounds grow.

Plan an A/B test

Simulates the next few weeks hundreds of times and runs three ways of splitting your traffic on each.

Conversion rate (%) Visitors per day Days available

How much better are the variants?

True lift of each variant (%) Relative to control, comma-separated: “5, 10” means 5.25% and 5.5%.

Test settings

Min. lift to detect (%) Significance α Power Confidence to call Simulations Posterior draws Top-two β Seed

“Auto” uses the smallest positive lift you entered (or half the ± range). Defaults are the Python tool's defaults.

strategy	lost conversions	picks the best	decided by the end	median day decided

Conversions lost so far, by day (average over simulations)

Same run as the command-line tool

The browser uses a different random number generator, so the simulated columns agree with the Python on average, not digit for digit. The sample size and test length are exact.

How it's built, results and limits: How it works →