283 replies to this topic

[ExoForce]

Stop posting this ideas. They will do exactly the opposite.

[Nightbird]

Cluster Fox, on 16 July 2020 - 09:03 AM, said:

Hey Nightbird, have you revised your Stomp criterion from the cumulative Std Normal when you plugged in actual player data?

While I get the same %chances as you for the same WL cumulative delta (Which I set between 200 and 1800), my random games using the actual player data (season 38 to 46) yields much less stomps than 35% and usually a smaller WL cumulative delta per randomly assembled team than your original simulation with fake T1 players.

Keeping the stomp gate at 47.5% only gives me about 27% of matches being stomps.

If I set the gate at (difference between random% & win% <= 37.5%) I get about 35% stomps.

However we've experienced that the PSR reset to 2500 increased the Stomp% significantly. The value after PSR reset might be around the 75% mark and I think that's quite conservative. Sadly Paul didn't give us stomp data, that would have been great to have, but maybe too telling.

There is too little data on stomps, so I tweak it until we get to what we know, around 30%, and use it to make a relative comparison. I actually got 35% stomps my first try with Jay Z's system, so I just tweaked until I reduced it to 30%.

[Nightbird]

Sweet zone has 54.7% of the 58% predicted at 120 games average, will we get there with the valve closing boost?

[Cluster Fox]

Just leaving this here...

Nice article on the compounding effects of win chance when dealing with two teams:

https://sabr.org/jou...-team-matchups/

Edited by Cluster Fox, 18 July 2020 - 03:33 PM.

[Nesutizale]

Cluster Fox, on 18 July 2020 - 03:33 PM, said:

Just leaving this here...

Nice article on the compounding effects of win chance when dealing with two teams:

https://sabr.org/jou...-team-matchups/

I have yet to read that completly and its to late for my attention so I will ask this. Does this paper work with fixed teams or a constant random team composition?

Asking because I wonder how teams of randoms compare to fixed teams. From just a quick overview it seams to work for fixed teams inside a fixed league.

So basicly in MWO when we say a teams of 12 wouldn't change for a season and allways play together and they all would only be from the same Tier with no matches between different Tiers as that would slightly need an adjustment of the W/L formula to predict outcomes.

Edited by Nesutizale, 18 July 2020 - 04:00 PM.

[Nightbird]

Nesutizale, on 18 July 2020 - 03:51 PM, said:

I have yet to read that completly and its to late for my attention so I will ask this. Does this paper work with fixed teams or a constant random team composition?

That article was for fixed team rosters. W-L for random teams is still accurate, it just takes longer to stabilize.

[Nesutizale]

How would that work? Would you take everyones personal W/L ratio, make an average value from that and then compare that to the other teams average W/L to predict the outcome?

While I can see the fixed teams W/L argument as true and 1:1 situations its hard for me to wrap my head around teams of randoms working the same way.

Edited by Nesutizale, 18 July 2020 - 04:05 PM.

[Nightbird]

Nesutizale, on 18 July 2020 - 04:04 PM, said:

How would that work? Would you take everyones personal W/L ratio, make an average value from that and then compare that to the other teams average W/L to predict the outcome?

While I can see the fixed teams W/L argument as true and 1:1 situations its hard for me to wrap my head around teams of randoms working the same way.

Picture 23 people on the two teams, other than yourself. On average across all matches, all 23 of those players will be average skilled. Just call this skill level A. If You < A, then your team will be You + 11A < 12A. This mean your W-L in the long run will be less than 1. If You > A, your team will be You + 11A > 12 A. Your W-L will be higher than 1.

By using WLR for the match maker, people with >1 WLR will be given harder matches, pushing their WLR down to 1. People with WLR < 1 will be given easier matches, pushing them up to 1 WLR.

This is not to say using WLR will result in a good match maker. It will still be a bad match maker. It just will be better than a terrible match maker.

[Nightbird]

55.7% / 58% predicted getting close

[Jman5]

I don't suppose it's possible you could convert WLR into Win Percent? TBH, I have never been a fan of WLR. The differences between a 1.2 and 1.3 feels like it should be larger, but in reality it's only about a 2% difference. It's also non-linear, which I just find kind of annoying. Finally, it's leading to this weird clustering in your graphs where you're bunching everyone with a 66.6% winrate and higher into a single column.

I don't think it's surprising at all that in a column representing 1/3 of all the possible win percentages you could have would see a spike even if the trend overall is downward?

The fact that you don't see a parallel spike in the low end makes sense to me too because indicates that there is only so much one can do to drag down a team of 12 people. Even if you tried teamkilling everyone, the game actively punishes and bans these individuals which filters them out.

[Nightbird]

Jman5, on 21 July 2020 - 04:58 AM, said:

I don't suppose it's possible you could convert WLR into Win Percent? TBH, I have never been a fan of WLR. The differences between a 1.2 and 1.3 feels like it should be larger, but in reality it's only about a 2% difference. It's also non-linear, which I just find kind of annoying. Finally, it's leading to this weird clustering in your graphs where you're bunching everyone with a 66.6% winrate and higher into a single column.

I don't think it's surprising at all that in a column representing 1/3 of all the possible win percentages you could have would see a spike even if the trend overall is downward?

The fact that you don't see a parallel spike in the low end makes sense to me too because indicates that there is only so much one can do to drag down a team of 12 people. Even if you tried teamkilling everyone, the game actively punishes and bans these individuals which filters them out.

Suppose you have 23 players with 0.5 WL (50%) and 1 player with 1.0 WL (100%). At the beginning of the match, 1 player disconnects. So you have 10 of those 0.5 WL players and the 1 WL player on a team for a total of 6 WL, against another team with 12 0.5 players also with a total of 6WL. Are the teams balanced? The 1 WL player is worth a lot more than 2 0.5WL players right?

The graph skews right so showing more bars >2 doesn't help much. I picked the range so that the peak is more or less center so it is easier to look at.

Edited by Nightbird, 21 July 2020 - 06:02 AM.

[Jman5]

Nightbird, on 21 July 2020 - 06:00 AM, said:

Suppose you have 23 players with 0.5 WL (50%) and 1 player with 1.0 WL (100%). At the beginning of the match, 1 player disconnects. So you have 10 of those 0.5 WL players and the 1 WL player on a team for a total of 6 WL, against another team with 12 0.5 players also with a total of 6WL. Are the teams balanced? The 1 WL player is worth a lot more than 2 0.5WL players right?

Perhaps he is, I couldn't say with certainty. Quantity has a quality all its own.

I just think the data would be more clear if the X axis values were all equally weighted and linear. That's why I like using win percentage over WLR.

Quote

The graph skews right so showing more bars >2 doesn't help much. I picked the range so that the peak is more or less center so it is easier to look at.

I don't doubt it skews right, but by how much and how does it look? For example 1.1% of the playerbase has 1.8-1.9 WLR, 1.9 to 2.0 also have 1.1%. How many players have 2.0-2.1 WLR? Does it suddenly spike up? I just think it would be more illustrative if the full 0-100% winrates were listed equally. If I had to guess you would see a sort of mini bell curve at the top end which probably indicates the effect high performing groups and an unstable PSR are having right now. But again, it's hard for me to know for sure when 1/3 of possible win percentages are lumped into a single column.

[Nightbird]

Jman5, on 21 July 2020 - 06:39 AM, said:

Perhaps he is, I couldn't say with certainty. Quantity has a quality all its own.

I just think the data would be more clear if the X axis values were all equally weighted and linear. That's why I like using win percentage over WLR.

Someone who wins every match in a match between average pugs is throwing a fox into a hen house.

It's easy to scrape the leaderboard data with free tools, feel free to do your own take on it.

Jman5, on 21 July 2020 - 06:39 AM, said:

I don't doubt it skews right, but by how much and how does it look? For example 1.1% of the playerbase has 1.8-1.9 WLR, 1.9 to 2.0 also have 1.1%. How many players have 2.0-2.1 WLR? Does it suddenly spike up? I just think it would be more illustrative if the full 0-100% winrates were listed equally. If I had to guess you would see a sort of mini bell curve at the top end which probably indicates the effect high performing groups and an unstable PSR are having right now. But again, it's hard for me to know for sure when 1/3 of possible win percentages are lumped into a single column.

It's just a long tail all the way to 10WLR and beyond. There's no meaningful information.

[Jman5]

Nightbird, on 21 July 2020 - 06:44 AM, said:

Someone who wins every match in a match between average pugs is throwing a fox into a hen house.

It's easy to scrape the leaderboard data with free tools, feel free to do your own take on it.

Would you mind telling me how? I guess you use Python, but I've never done it before and don't know where to start. I've always relied on looking at spreadsheets other people have provided, or manually copy/pasting limited dataset.

Quote

It's just a long tail all the way to 10WLR and beyond. There's no meaningful information.

Yeah I understand why you did that using WLR on your X axis, but that's exactly why I don't like using WLR.

[Nightbird]

Jman5, on 21 July 2020 - 07:28 AM, said:

Would you mind telling me how? I guess you use Python, but I've never done it before and don't know where to start. I've always relied on looking at spreadsheets other people have provided, or manually copy/pasting limited dataset.

I use the chrome extension Web Scraper. After you install it, there is a tutorial, I watched the first and easiest one and it on the leaderboards.

[Nightbird]

56.7 of 58

[Nightbird]

57.8 / 58.1

[Nightbird]

We're at the average games=120 milestone which corresponds to the top right predicted graph. Between the 0.8 - 1.2 bars, it was predicted that 58.1% of the pop will be within it.



In actuality we see 58.4%.



Going forward I would like to examine the 0.9-1.1 range instead. Since this represents a tighter grouping. At 120 average games, 39.7% of the pop was predicted to be in this range and 38.7% was the actual. The WLR system predicted that 65.6% of the pop would be in this range, a 1.65x improvement.

At 240 games, the next milestone, it is predicted that 44.6% of the pop will be in this zone. I'll only update on a weekly basis since it'll take a while to get there.

[Nesutizale]

What is the point of that graph? I didn't follow the topic that much. Is it to show that W/L is better, similar or something else to the current system?

Edited by Nesutizale, 25 July 2020 - 03:35 PM.

[Nightbird]

Nesutizale, on 25 July 2020 - 03:35 PM, said:

What is the point of that graph? I didn't follow the topic that much. Is it to show that W/L is better, similar or something else to the current system?

It shows that simulation can accurately predict how a PSR system would work without having waste months trying it out. You can also compare different systems to see which would work better.