User blog comment:Sirebel/Mathematician Required - Apply Within/@comment-27170954-20170523122715

I won the car in question (NISSAN Skyline 2000 GT-R (KPGC10)) and fully upgrded the car through 12 different paths. I collected the PR and upgrade content for 269 unique upgrade configurations.

running the regression i come up with the following 8 values for the base PR and each upgrade category.

Using these values to estimate PR based on the upgrade configuration and rounding to 1 decimal place yields very good, but not perfect results. 2 of the 269 predictions are still off by 0.1.

My dataset is not perfect. I tried to avoid duplication, but once i got past 19 of 26 upgrades i couldn't avoid it. As a result, some of the upgrades at the top of each brach are under represented (E4, D3, etc) while some are over represented (S3, Ex3, TW3). This is the result of bad planning for the upgrade paths. For example, there are 28 possible configurations with -2 below FU, but i only got 9 of them because of my duplication issues.

I still believe, even after doing this, that it is possible to get 0 errors with the right dataset. I'm working on a set of upgrade paths that allows me to collect all 28 -2 configurations along with a larger dataset. This will take me a little while, as planning the (about) 12 FU paths with no duplication is time consuming, and i don't get a lot of free time these days.

I as can't yet convincingly say with any statistical precision each level in an upgrade branch makes the same PR contribution. I belive it to be the case, but the data i collected was insufficient to support that conclusion.

I report back in a couple weeks with what i learn (it took me three weeks to get this far)