Offline replay · runs in your browser

The lab

Seven bandits replay the same synthetic click log. A round only counts when an algorithm picks the arm that was actually logged, and its feedback may arrive late or never. Change the arms, the delays or the contestants, then pull the lever.

Ready.

estimate ≈ 7.0 s

Settings

Quick set-ups

Start from a scenario

Algorithms

Contestants

7/7

10 arms · 300,000 events

The logged data

A uniformly random logging policy showed one of ten items per event; each click happened with that arm's rate. Drag a fader to change an arm's click-through rate.

.11

.29

.51

.13

.18

.26

.01

.38

.30

.06

Ringed arm = best arm (μ* = 0.509). Same per-arm click rates as the 2023 course log (best arm 2 at 0.51).

Log seed

Seeds the synthetic log (numpy recipe: integers, then random).

When do clicks arrive?

Delayed feedback

Each event's feedback arrives after round(Pareto(α)) rounds. Small α = heavy tail: with α = 0.2 a third of the feedback takes over 250 rounds and one in seven takes longer than 20,000.

1.00

0.20

1.00

0.20

1.00

Horizon, repeats, seed

Run length

Matched rounds per runT10,000

The coursework evaluated 20,000 matched rounds.

Repeats averaged5

Each repeat is a fresh bandit drawing from the same generator, like the notebook's 10-repeat plots.

Algorithm seed

default_rng(seed) per algorithm; 90051 is the notebook's seed.

Cumulative average reward

Running click-through rate over matched rounds, averaged over completed repeats, as in the notebook's plots. Higher is better; the dashed line is the best arm.

Pull the lever to start a race. Curves appear as soon as the first rounds are matched.

Standings

The table fills in as soon as a race starts.

Heatmaps appear once each algorithm completes its first repeat.