Weighted Random Picker

Selects model randomly, where the probability of a model being selected is proportional to its score.

It is registered as type weighted-random-picker and runs as a modelselector picker.

What it does

1. Receives a list of ScoredModel candidates.
2. If all candidates have a score of zero or less, it delegates to the random-picker for uniform selection.
3. Uses the A-Res (Algorithm for Reservoir Sampling) algorithm to perform mathematically correct weighted random sampling.
   • Generates a random key for each endpoint based on its score: $key_i = U_i^{(1/w_i)}$ where $U_i$ is a random number in $(0,1)$ and $w_i$ is the endpoint's score.
   • Selects the candidates with the largest keys.
4. Returns the top candidate.

Behavioral Intent

This picker resolves the trade-off between max-score-picker and random-picker. It prefers higher-scoring model while maintaining exploration and avoiding extreme hot-spotting.

Inputs consumed

• Consumes the list of ScoredModel results and utilizes the Score value.

Configuration

The plugin has no config.