Adjust topic initial regret initialization

[guilherme-brandao]

Purpose of Changes and their Description

PROBLEM

• To calculate the initial topic regret, we need to have at least one participant with enough experience to be used in the StdDev.

• workerRegrets are selected based on the worker experience:
  experienceCount > 1/alphaRegret

• After the last migration, we deleted all regrets, set the initial regret of topics to 0, and also set all scores to 0. All participants were reinitialized.

• A high number of workers with equal scores means high volatility in the active set and, consequently, a low chance of experienced participants in the active set. The problem gets a lot worse for topics that have long epochs.

SOLUTION

• If there are fewer than 10 workers (or not enough to calculate a reliable standard deviation), we include regrets with minimal experience.
• In cases where regrets are used with insufficient experience, set the dummy regret to the 25th percentile of existing regrets, without adjusting by a standard deviation multiple.
• Define "not enough workers" as fewer than 10 workers.
• This approach activates the intended functionality once there are at least 10 workers, each with at least
  1/alphaRegret epochs of experience. If these conditions aren’t met, we’re in a cold start, prioritizing rapid inclusion to avoid gaps.

The essence of the change is to maintain the purpose of regret equilibrium before calculating standard deviation, aligning with the design intention of regret handling.

Link(s) to Ticket(s) or Issue(s) resolved by this PR

https://linear.app/alloralabs/issue/PROTO-2806/investigate-potential-inference-synthesis-issues-found-by-research

Are these changes tested and documented?

• [OK] If tested, please describe how. If not, why tests are not needed.
• [OK] If documented, please describe where. If not, describe why docs are not needed.
• [OK ] Added to Unreleased section of CHANGELOG.md?

[relyt29]

Conceptually looks good. Minor comments

[amimart]

Nice, just noted some simplification

[kpeluso]

The PR description should be updated to reflect the latest intentions implemented by this PR / description of what's actually here @guilherme-brandao

[kpeluso]

@guilherme-brandao the PR template was not fully filled out

[kpeluso]

Just 1 missing param. Otherwise, matches this description

[github-actions]

The latest Buf updates on your PR. Results from workflow Buf Linter / buf (pull_request).

Build	Format	Lint	Breaking	Updated (UTC)
✅ passed	✅ passed	✅ passed	⏩ skipped	Dec 3, 2024, 7:37 AM