There has been some spirited discussion on how CSR is divvied up after a game.
Do kills matter? And if so is it K/D, KDA, or K/min?
Does playing the objective harm your CSR progression?
So… I’ve started collecting data. At the end of each game I’m going to be grabbing player’s K, D, A, score, and CSR change (while waiting for the next game to start). Throw them in a spread sheet and see what correlates with what?
What I’ve done is use Sheets to calculate out the player’s K/D, KDA, and K/min. That’s easy. And estimate their objective score by subtracting kills and assists from their total score. The CSR change is estimating the red or blue change on the post game screen - which is now the game just played!
I’m then using Sheet’s CORREL function to look for correlation on CSR change with K, D, K/D, KDA, K/min, Score, and Objective Score.
Nothing so far… but very low sample size. I’ll post again when I get a couple of hundred results into the sheet.
Any suggestions welcome before I go too far. Math’s isn’t my forte.
It takes into consideration ever individual players performance per game. Not just your own.
It also creates a CSR for the two teams. And calculates a win/ loss probability that is also taken into consideration.
There is also a variable in play that decreases lost CSR or gained CSR if someone on your team or the enemy team leaves the match.
So win or lose a ranked game. These are the variables that are considered in regards to how much reward/ rank is obtained.
Player 1 CSR
Player 2 CSR
Player 3 CSR
Player 4 CSR
Team CSR
Team Win/ Loss Probability.
Player 1 Performance
Player 2 Performance
Player 3 Performance
Player 4 Performance
Player 1 Match complete
Player 2 Match complete
Player 3 Match complete.
Player 4 Match complete.
Team Win/ Loss
Player 1 KDA
Player 2 KDA
Player 3 KDA
Player 4 KDA
I tried reading a discussion on the maths involved in TrueSkill. It was like wow. Multiple passes at comparing individuals. It was way over my head.
Yep.
But the argument is that people are convinced that there is a clear correlation. eg. K/D and CSR gains.
Simple data collection should at least show if these anecdotal claims are true or not. But I’ll probably need a fairly large sample size to reduce the noise from those other variables.
Just use the data on halo tracker. Use the leaderboards for the ranked playlist and look at all the top 100 players stats vs the bottom
100 players stats.
Seriously low sample size so far… like barely a page of results.
But no correlation on K, K/min, or Total Score.
Low correlation on D, K/D, and KDA.
Moderate (negative) correlation on Objective Score. And only just, like -0.505. But this is one of the big ones to watch. Especially as Total Score appears to be poorly correlated.
I’ll go hard over the next week getting more data…
Cool idea but pretty impossible to determine too much on a game by game basis I think. Players will have different MMRs going into the game and you could have players whose losses or gains are completely tied to that.
You yourself are losing little CSR at the moment as your MMR is higher. So your stats will have less impact on your immediate CSR changes if your MMR remains high.
We also must consider the player expectation Vs performance. Every player is not expected to have a flat performance against each other player. The highest MMR player is expected to have the best performance and the lowest, the worst. A 1.2 k/d will mean different things to the system depending on which player achieves it per game.
MMMR/CSR is a whole mess and the designing of it is crazy. All I know is, I consistently get pretty close and good games in ranked. Like never before.
My guess would be the system is pretty similar to what was used in halo 5. On the old waypoint Menke went into good detail about how the ranking system worked.
My take away is essentially the system has a hidden MMR and a visual CSR. The reason for this is MMR is very much so performance based so there can be wild swings match to match, whereas CSR increases on decreases in a more fixed way. I do not know the exact values in halo infinite but in halo 5 it took into account if your team won or lost. If you won your CSR always went up, even if just by 1pt. And if you lost your CSR either dropped or had a 0 if you performed well but still lost.
CSR in halo 5 was based on converging to your MMR and you would get big gains until they converged because they wanted to give a sense of progression so your initial CSR after placements was lower than what MMR was.
As for how MMR was calculated, my takeaway is the game looked at did you win, what your kills per minute was and what your deaths per minute was. Each of these variables seemed to be assigned a weight with win being most important then kills/min and then deaths/min. After a few games the system gave a prediction and if you outperformed the systems prediction your MMR would go up and if you underperformed MMR dropped.
The system also took into account the competition you were playing against and so it knew the range for how a typical onyx player would perform as opposed to someone who was gold rank. So, for example if an onyx player had a silver they were matchmaking with and then the game filled in the other two teammates with plats and they played a team of plats, the system would expect the onyx player the heavily slay for their team.
From what I can tell with my playtime in infinite, the ranking system seems to be very similar to this.
As a side note, Menke stated other stats like objective points, assist, etc. were tested and did not improve the accuracy of predicting performance of individual players and win outcomes and slowed the processing time. Therefore only W/L, KPM and DPM were used.
I really miss Menke and his participation on Waypoint.
Pretty much.
Your MMR is a curve, not a number, so you need some sort of CSR to represent it.
Plus, as you mentioned, you can use CSR to reward wins which would otherwise not earn MMR (eg. beating a side ranked below you). The danger is that if the player doesn’t eventually prove they deserve this rank it’s taken away (on the next loss) - and this is what seems to frustrate people the most.
Your CSR is based around the left hand edge of your MMR curve. Three standard deviations below the mean to be precise. As the system gains confidence in your rank (consistent performance) your MMR curve narrows and gets taller. Eventually your standard deviation is small enough that the CSR and mean MMR are pretty much the same.
Yep. The system uses all the MMR curves - averages them together and then subtracts them to work out the predicted winner. The result is then used to compare all the MMR curves together (apparently this takes multiple passes) and to decide how much MMR to take or add to each curve. If you perform close to expected then your curve is a pushed towards narrowing.
Don’t know if you’ve read this stuff; https://www.moserware.com/assets/computing-your-skill/The%20Math%20Behind%20TrueSkill.pdf
It’s not clear what factors or weightings are used. Except for K/min - but I think the bar for this is set pretty high (you need to be really dominating). I also suspect (personal opinion) that placement increases the performance weightings over W/L…
And this pretty much mirrors what was said in the TrueSkill2 discussion paper by Microsoft.
Except the bit about DPM. Do you remember anything else Menke may have said about that?
I have read the link you added, but it was a while ago so I’ll likely read over it again on my lunch break as I find this kind of stuff fascinating.
As far as other things Menke mentioned, it primarily was around player base/population and matchmaking times. Wanting the system to still get close matching with chances of winning for either team being as close to 50/50 as they can get it so that is why a lot of high skill players feel like they have to hard carry a lot. There are not enough of them queuing at the same time so the matches are balanced to be close overall but the individual skill of each player may be wider than what you would want to see in a ranked match. Factor on top of that trying to get relatively quick queue times and a good/stable connection and you get what we have now.
I did peek at his twitter recently and it seems he has since modified and advanced trueskill2 since leaving 343 and going to riot games but won’t get into specifics but that the basic principles are similar. So that also explains why infinite ranked system feel slightly different (and worse imo) than halo 5 did. They likely also tweaked things and are trying to find the right balance for everything and every type of player.
Why do you feel that way? His transparency was unparalleled by anything I have ever seen from any first person shooter dev and the matches in halo 5 even with the limited population were very close. On top of that he truly took the time out of his own day to interact with the community.
His matchmaking system has ruined matchmaking for many people, especially friends who like to play together and aren’t exactly the same skill. This is one of the things that made Halo popular… But not anymore.
Were there people that abused the system? Sure. But for those of us that didn’t, we are now stuck losing 85%.
And guess what? There are atill just as many people abusing the system, if not more now that it is Free-To-Play and people can make as many Smurf accounts as they want…
So I’m not sure what you are trying to accomplish with your comment.
You have played 60 games with Primus Ego Sum in ranked and you have a 50% win ratio with them. Over 202 games you have won 50.5% with them. Any other player you have played with more than 5 times you have a 60% win ratio with lmao. What is this??
Back in the day if you played with a better friend on Halo 3 you would get stomped into a puddle, because the game would only do you a favour of making everyone equal to or below your friends rank, but say you’re a 10 and they’re a 40, you are playing minimum level 30s and above. You’re getting destroyed. In social parties would win way too often. Now it’s more fair. Your stats are proof of that.
I think it’s unlikely you will reach sufficient sample size with one player’s set of games to get a good correlation.
However, I would argue that is a meaningful result in and of itself. Populations don’t play the game, players do. Over a reasonable sample of games (30?) if you can’t draw a correlation between your performance (individual score/KPM/win-loss) and your CSR change (even allowing for your performance being “raw” e.g., not adjusted for expected performance), then it’s easy to understand why people are often frustrated by the system.
It can be good at producing matchups and make statistical “sense” at the population level and still not be fun at the individual level, which I suspect is what is occurring here.
If you really wanted to geek out you could probably create crude rules to account for some expected performance. You could use HaloTracker to pull expected win percentages or create some rules around individual expected performance (e.g., ratio of Player CSR to average team CSR) and use those to adjust your individual performance metrics before feeding them into the correlation.
Exactly.
CSR has a graph per player on halo tracker and you can correlate gains and losses per game.
Whatever result this guy gets is solely based on his own stats and does not consider any of the aforementioned stats needed to consider to give an accurate result.
I tried explained and pointing in the right direction but it fell on deaf ears/ eyes I’m afraid. He seems to be set on doing things his way to get to a result he wants to validate his own argument in another thread.
That’s bias and not the truth of the matter.
So I’ve lost interest in offering advice.
