[00] Structures of 'Role Star Power'

Continued on Intro post, I will explain the details.

 

Our ANN learns general past data of players (2019, 2020) and it especially learns that stats of 'Role Stars' are good. Then, what is 'good'? Think of people who can cast a vote. I don't know exactly but I assume that voter would think like "I chose Super as he is the best aggressive Rein!" not "I would go for Super as he is ranked 1st in Final Blows and 2nd in Shatter Kills".

 

Why do we think of Super when it comes to great Main Tank? Are there special reasons?

 

So our AI would learn that "Super is good, right, so Final Blows and Shatter Kills which Super is good at should be weighted for Main Tank!" not 'Final Blows & Shatter Kills are significant roles for Main Tank, oh then, Super is good!". You should mind the utter difference between the two.

 

With my PR philosophy, I will now show who is really acting like past 'Role Stars' in 2021 season through 'Role Star Power' (==RSP). The more point a player gets, the more probability he/she would have to be nominated as Role Star.

 

Caution - Like I said, the RSP does not to be perfect, it tries to be natural human (or Role Stars Voters). Therefore, getting high rank does not assure the higher RSP player is better than others.

 

In a nutshell, ANN drank a whole lotta beer and tries to rank OW players like many of us usually do.

 

 

2021 Model

Training data - Collected through 2019 and 2020 season when role star index is available.

 

Role Star Pred. Model - It predicts '2021 Role Stars' with variables below.

 

Explanatory Variables

Basic Variables [2021 Ver.]

Final Blows - A value that is needed to get all the spotlights. {Dealing like 5 damages with Moria and taking final blows would make your Nanoed-Genji angry but still it is really significant signal in the game}

 

Eliminations - 1st important political resource in the competitive OW. {You won't be blamed for a defeat with Gold Elim medal}

 

Hero Damage - 2nd important political resource in the competitive. {Zarya and Moria users usually trash talk with the value}

 

Solo Kills - This is for one-shot snipers at first glance. But Projectiles, Tanks and even Off Spt. can show their supremacy with the stat. {But when Lucio or Mercy is doing bunch of solo kills, you may as well give up the game}

 

First Blood Rate (FBR) - Good players in a bad team can be compensated for with the stat. First Blood is essential in winning the teamfight, but bad team still loses. {Still you are not justified for an initiative Rein Charge or 24/7 flanking Hog}

 

First Blooded Rate (FBdR) - Bad players in a good team can be penalized with the stat reversely. {You are also not justified for playing hide & seek with 11 players}

 

Winner - Only for Tournament winners (Grand Finals not included as Role Stars are selected before post-season). It may be the greatest impact to become Role Stars to be honest. {Carpe achieved Role Star tho}

 

POTM - Another great index of 'Impact'. POTMs are not always perfect, but it usually does!

 

Env. Kills - Only for Main Spt & Main Tank players. With their innate dependency, they can hardly show their impact. Env Kills fortunately display their ability.

 

Ultimates Negated (as Sigma & D.Va) - Only for Off Tank players. Measuring how good they are is always hard to figure out. Not only do they have to push the frontline, they also have responsibility to care the backline. With most variables implying the offensive behavior, this one will balance them to some extent.

 

https://overwatchleague.com/en-us/statslab Statslab offers the open data

 

Power Computation

Like Role Stars are voted by the many, I will invite multiple AI robots {can't think of better names}. Total 33 AI friends would learn the data. It's the same data, but the data order they receive is not the same. So they will answer similar but different outputs. By averaging their outputs, the model can finally printf() the Role Star Power!

{Anyways, I chose the number 33 to pray for the next 33th hero}

 

Drawbacks

Those are the variables chosen by myself within public OWL data. So there is a high chance that undisclosed data would have meaningful Xs or I haven't discovered significant ones in the open data. Whatever the thing is, the current model seems pretty okay. But I'll try to update (2021.07.24) variables whenever possible.