Quick Pick Survey asks people to choose between two items at a time, then ranks the list from the answers. Which two items to show next is chosen by an information score: for every pair in the pool, how much would one answer shrink the uncertainty of both ratings? At 500 items that is about 125,000 pairs per answer, each with an exp, a hypot and a square root.
The simulations behind the selector's design play thousands of answers. The longest, 500 items and 4,000 answers, took 34 seconds for a single seeded run.
Most pairs cannot win
We only need the best pair, so we do not need the exact score of a pair that cannot beat the best one found so far. Two facts give a cheap ceiling:
- An answer's information peaks when it is a coin toss: p(1 - p) is never above 1/4. That alone bounds what each item can gain, whoever it meets.
- For pairs far apart in rating, p(1 - p) = 1 / (2 + 2 cosh x), and cosh x >= 1 + x^2/2, so it is at most 1 / (4 + x^2). That needs no
exp.
If a pair's ceiling is below the best score so far, we skip it. The ceiling has a margin far larger than any float rounding, so a skipped pair could never have won.
Same picks, not just close ones
The selector breaks ties with a tiny random factor, one random draw per pair. A skipped pair still takes its draw, so the random stream lines up and every later pick is the same. The test runs seeded simulated surveys through both the new loop and the original one, records the random draws, and checks that every pick matches and that both draw exactly as many. A deliberately loose ceiling turns seven of those tests red. The design numbers that come out of the simulation suite are byte-identical before and after.
What it bought
- The simulation suite: 77 s to 33 s on 4 cores. The 4,000-answer cell went from 59 s to 19 s.
- The live "next pair" call, same function: 9.6 to 3.2 ms at 500 items, 2.8 to 0.8 ms at 220.
The run's notes have the profile and what to try next.