New sub-OPR Calculations

[Ed Law]

Hi,

Sorry it took so long. I had a very busy summer. I recalculated all the autonomous OPR, teleop OPR and climb OPR to take into account the surrogate matches that messed up the data. It was discussed in the thread http://www.chiefdelphi.com/forums/sh...d.php?t=116791. When I have more time, I will explain the math behind what I did.

You can download the latest database at http://www.chiefdelphi.com/media/papers/2174

I hope it is not too late for teams going to IRI. Good luck to everybody going to IRI.

[Basel A]

It seems like this post explains what you did pretty well (at least for anyone who understands the base OPR calculations). Or were there any other twists?

[Kevin Leonard]

Something interesting must have happened to Archimedes rank #31, FRC 4753.
Auto OPR of 46.8?

[Ed Law]

Originally Posted by Basel A

It seems like this post explains what you did pretty well (at least for anyone who understands the base OPR calculations). Or were there any other twists?

Yes, that is how I did it. I didn't like scaling the score b based on number of matches. It has no mathematical basis. What I proposed to do was just an intuition. It seems more logical but I did not do the math at that time. I finally did the math and what I propsoed was the best way to do it.

However, as Kevin pointed out, it created weird numbers in some rare situations. It happens when teams have negative sub-OPR numbers and especially when teams were involved in surrogate matches or noshow or DQ matches. The reason is I scaled each of the sub-OPR numbers to match the overall OPR number.

I published a new set of data. Instead of scaling, I shift each sub-OPR numbers proportional to the absolute value of the sub-OPR number of the difference that needs to be made up. It works really well. If all sub-OPR numbers are positive, it gives the same number as by simple scaling.

I don't have time right now to explain in more detail. I will try to do that this weekend. I only spotcheck some events. If somebody finds something weird, please let me know.

[Ed Law]

The purpose of this post is to explain exactly how to calculate the sub-OPR, i.e. autonomous OPR, teleop OPR, climb OPR to adjust for issues with surrogate matches, no shows and DQ matches. I would like to ensure all programs that calculate OPRs will give identical results so users can trust the data.

Ether explained the issue very well at http://www.chiefdelphi.com/forums/sh...1&postcount=22 so I am not going to repeat it.

It is fairly easy to find out what the missing scores are whether it is due to surrogate match, no show or DQs. If you add up all the scores from all of the team's matches, and then subtract the AP, CP and TP from the Team Standings data, you will get the missing score.

When solving for A x = b, imagine b is the missing score vector, we can solve for x which represents how much effect it has on each of the teams due to the missing score.

Hence it is possible to scale the score vector b before calculating the OPR which is x in the equation.

Considering the fact that if A x1 = b1, A x2 = b2, then it is true that A (x1+x2) = b1+b2. In my program, it is much easier to scale the x than to scale the b, so that's what I did.

Here are the codes I used shown in pseudo code

shift = OPR - (OPR_auto + OPR_tele + OPR_climb)
sum = abs(OPR_auto) + abs(OPR_tele) + abs(OPR_climb)
OPR_auto = OPR_auto + shift * abs(OPR_auto) / sum
OPR_tele = OPR_tele + shift * abs(OPR_tele) / sum
OPR_climb = OPR_climb + shift * abs(OPR_climb) / sum

Example: Finger Lakes Team 4023

All teams were schedueld to play 9 matches. Team 4023's record was 1-6 which means they missed two matches. The alliance in those two matches scored a total of 92 points which is missing from the calculation of Team 4023's sub-OPR score. It makes their sub-OPR number even worse. The side effect is it also affect all the other team's sub-OPR numbers.

Using old way of calculating sub-OPR,

OPR_auto = -3.9
OPR_climb = -8.3
OPR_tele = 0.6

Using the new proposed way of calculating sub-OPR,

OPR_auto = 0.0
OPR_climb = -0.1
OPR_tele = 1.3

This matches the overall OPR of 1.2 much better and a better estimate of what actually happened than a OPR_climb of -8.3. I also compared a few other cases and I am very happy with the results but I am not listing them here. Check it out and let me know what you think.