1868011 - auto-adjust bisection point bias based on the good/bad tendency

Status: UNCONFIRMED

(Testing :: mozregression, enhancement)

Description

[tophf]

Steps to reproduce:

When using the current neutral bias of always choosing the middle of a range the number of steps generally doesn't depend on the bad/good choice.

Actual results:

The number of steps doesn't change regardless of how close the result is to an edge of the current range, here's a mock of 100-200 range bisection:

1. build 150 is good
2. build 175 is bad
3. build 168 is bad
4. build 159 is bad
5. build 155 is bad
6. build 153 is bad
7. build 152 is bad
8. build 151 is good

Expected results:

Use the last 3 user decisions for good/bad to calculate the bisection point bias, for example:

maxInfluence = 0.8
bias = ((last3 is good?0.5:-0.5) + (last2 is good?0.75:-0.75) + (last is good?1:-1)) / (0.5 + 0.75 + 1)
build = bad + (good - bad) * (bias * maxInfluence + 1 - maxInfluence)

Comment 1

[tophf]

My formula example is wrong, apparently, here's a fixed one:

build = bad + (good - bad) * (0.5 +  0.5 * bias * maxInfluence)

But maybe there's an entirely different better one.