compere

Glossary

Glossary of terms

The vocabulary of pairwise-comparison ranking, defined in plain language.

Pairwise comparison
A single judgement between two items: “which is better, A or B?”. compere aggregates many of these into a ranking rather than asking for absolute scores.
Multi-armed bandit
A framework for choosing among options under uncertainty, balancing exploiting known-good choices against exploring uncertain ones. compere treats “which pair to compare next” as a bandit problem.
UCB1
Upper Confidence Bound — the bandit rule compere uses for pair selection. It scores each entity by observed win rate plus an exploration bonus that shrinks as the entity is compared more.
Exploration constant (c)
The weight on the UCB exploration bonus. Higher c means more exploration of under-compared entities. Default 1.414 (UCB_EXPLORATION_CONSTANT).
Unexplored weight
A large initial priority (default 1000.0, UCB_UNEXPLORED_WEIGHT) given to entities with no comparisons yet, so new items are surveyed before they are scored.
Elo rating
A rating system from chess. Each entity has a numeric rating; after a comparison, the winner gains and the loser loses points, scaled by how surprising the result was.
K-factor
The step size of an Elo update. A larger K makes ratings move faster (and noisier). Default 32 (ELO_K_FACTOR).
Initial rating
The Elo rating a new entity starts with before any comparisons. Default 1500 (ELO_INITIAL_RATING).
Expected score
The Elo-predicted probability that one entity beats another, computed as 1 / (1 + 10^((opponent − rating) / 400)). The rating update is proportional to actual minus expected.
Similarity pairing
An alternative selection strategy (compere/modules/similarity.py, exposed at /comparisons/next) that pairs similar entities instead of UCB-optimal ones.
Leaderboard
The list of entities sorted by Elo rating, returned by GET /ratings. It is a deterministic function of the comparisons recorded so far.
Stopping rule
A criterion for deciding you have collected enough votes — for example, when the top of the leaderboard stops changing across new comparisons.

Want the maths behind these? See how compere works.