Hey players,

this is more of a chat and sharing session than anything else, so please treat it as such.

To recap some of the things I have shared in the past.

We actively calculate each players rank based on the result of every match, and we lovingly call that rank "Mu". Before the first battle, the player is given a rank of 25. That rank is updated after the battle, based on the weighted average of the persons duration in the battle and whether he won or lost. Simplified (the calculation is considerably more complex)

NewMu=OldMu+"Player seconds in Battle/Battle Duration" * Win/Loss

So if you lose, your Mu goes down, and if you win, your Mu goes up.

This is a modified version of elo ranking, used in many competitive sports such as chess, major league team sports and esports. Eventually your Mu will converge and stabilize around your "true skill", which is where you will win and lose equally against either players with the same Mu. However, since this is a team game, convergence will happen slower and you may find that even having the best game of your life, will not influence the match enough to secure a win. I will, however, demonstrate that it works very well to predict player skill.

The underlying problem is that after one match, everyone is very close to 25, and 2 battles, even 10 battles in, only the very best (and worst) have begun to be different from the pack.

That's why it's imperative to find a proxy for Mu for the first battles, which is what comes next in our findings.

Now to the data. I have been working with a big sample of player data, testing multiple theories, some from the forums (WP/Death), the ever classic K/D of course and a time based WP/second and the results are very promising.

DATA

By creating buckets of Mu, I can calculate the three ratios of the players within each those Mu buckets and create a correlation table against Mu. That Correlation is then shown graphically on the top Chart.

Seeing that two of them are obviously logarithmic in nature, I normalize with the log function and get the bottom chart, Normalized Correlation.

Calculating and also just analyzing the graphs, we find that the correlation between WP/s and Mu is a towering 99%, and a bit lower for the other two, but still statistically very relevant. It basically means that all of them could be used in place of Mu in the beginning while Mu converges, and even in place of Mu overall.

Now, our next step is to implement a better use of this data. One simple way would be to say, instead of exiting the Academy at an earned WP basis, it's not until you actually reach a minimum threshold of WP/s. It is also imperative to utilize this information more during the teambuilding part of the matchmaker. In any sport, if the two best players are on the opposite side, everyone, even the bad players, can have fun and be inspired by the good players. If both of them are on the same side, nobody has fun.

I hope you enjoyed this little insight piece

P.S.
Those with eagle eyes will notice a weird anomaly in the two lowest Mu brackets for the both WP/Death and K/D ratios, but not for WP/s. My theory, is, and not based on prejudice at all, is redline snipers. My reasoning is that they are able to avoid death rather easily, they will be able to pick off stragglers and low hitpoint suits on a regular basis but sadly, have little to no relevance to the battle result, as they do not hack nor defend objectives effectively. Why their WP/s does not show that, I theorize, is because they spend quite some time getting to a mountain top, and or with a dropship to a tower, and if they die, they are forced to do so again. Feel free to burn me at the stake, and/or voice your alternative theories.

This is awesome. Thanks for sharing this!

How is each team made up? In other words, say you have 32 players with similar Mu scores, how does it determine which 16 players are on which sides. I always scratch my head when Scotty puts 2 full squads of talented players against 16 randoms. Surely it would make for a better match if those squads were on opposite teams, no?

That's the aspect we desperately need to fix.

Hey players,

this is more of a chat and sharing session than anything else, so please treat it as such.

To recap some of the things I have shared in the past.

We actively calculate each players rank based on the result of every match, and we lovingly call that rank "Mu". Before the first battle, the player is given a rank of 25. That rank is updated after the battle, based on the weighted average of the persons duration in the battle and whether he won or lost. Simplified (the calculation is considerably more complex)

NewMu=OldMu+"Player seconds in Battle/Battle Duration" * Win/Loss

So if you lose, your Mu goes down, and if you win, your Mu goes up.

This is a modified version of elo ranking, used in many competitive sports such as chess, major league team sports and esports. Eventually your Mu will converge and stabilize around your "true skill", which is where you will win and lose equally against either players with the same Mu. However, since this is a team game, convergence will happen slower and you may find that even having the best game of your life, will not influence the match enough to secure a win. I will, however, demonstrate that it works very well to predict player skill.

The underlying problem is that after one match, everyone is very close to 25, and 2 battles, even 10 battles in, only the very best (and worst) have begun to be different from the pack.

That's why it's imperative to find a proxy for Mu for the first battles, which is what comes next in our findings.

Now to the data. I have been working with a big sample of player data, testing multiple theories, some from the forums (WP/Death), the ever classic K/D of course and a time based WP/second and the results are very promising.

DATA

By creating buckets of Mu, I can calculate the three ratios of the players within each those Mu buckets and create a correlation table against Mu. That Correlation is then shown graphically on the top Chart.

Seeing that two of them are obviously logarithmic in nature, I normalize with the log function and get the bottom chart, Normalized Correlation.

Calculating and also just analyzing the graphs, we find that the correlation between WP/s and Mu is a towering 99%, and a bit lower for the other two, but still statistically very relevant. It basically means that all of them could be used in place of Mu in the beginning while Mu converges, and even in place of Mu overall.

Now, our next step is to implement a better use of this data. One simple way would be to say, instead of exiting the Academy at an earned WP basis, it's not until you actually reach a minimum threshold of WP/s. It is also imperative to utilize this information more during the teambuilding part of the matchmaker. In any sport, if the two best players are on the opposite side, everyone, even the bad players, can have fun and be inspired by the good players. If both of them are on the same side, nobody has fun.

I hope you enjoyed this little insight piece

P.S.
Those with eagle eyes will notice a weird anomaly in the two lowest Mu brackets for the both WP/Death and K/D ratios, but not for WP/s. My theory, is, and not based on prejudice at all, is redline snipers. My reasoning is that they are able to avoid death rather easily, they will be able to pick off stragglers and low hitpoint suits on a regular basis but sadly, have little to no relevance to the battle result, as they do not hack nor defend objectives effectively. Why their WP/s does not show that, I theorize, is because they spend quite some time getting to a mountain top, and or with a dropship to a tower, and if they die, they are forced to do so again. Feel free to burn me at the stake, and/or voice your alternative theories.

Even a squad of six randoms will consistently beat six unsquadded randoms.

I strongly believe that squad members should get additional mu, that grows more as the squad size grows (i.e. more than linearly)

One approach to consider is using social networking algorithms. Are these players frequently squadding together? If so, then they're likely going to be well-coordinated, that should give them a higher weight for deciding teams, as opposed to people who randomly found each other using the squad finder. They would have a much lower frequency of squadding together in their histories, and will likely be less effective as a unit with a correspondingly lower weight for deciding team balance.

Edit:
Also, at the end-of-match screen are players significantly above or below the median WP score for the match (say by a factor greater than 3)? In these cases it might make sense to make a larger adjustment to their Mu scores.

That's the aspect we desperately need to fix.

I disagree with your first point and probably think the second point be bracketed for the moment.

First, I think many players who squad together will have inflated Mu scores. It is much more difficult to run solo. If one or two players squad constantly with two other players who are much better than them, then the prior will appear to be better than they are. I, for one, used to squad with people who had much more SP than I did (still do, in fact) yet was able to use them to score much more warpoints than I would have if I were alone. Would have been a rude awakening when I squadded alone after that. I think such experiences are essential to FPSes.

Second, I think that we should implement a ranking algorithm before adding certain provisions to it (such as, a certain player's score in relation to those of other players in a given match implies that this player ought to receive x-times Mu points for it). To me, adding a ranking algorithm at all to Dust will better the experience of the game so much that whether it is rudimentary or not means less than whether it is introduced ASAP.

Just to formalize what I was spitballing earlier. HereGÇÖs one algorithm for approaching this problem:

Use a second variable (IGÇÖm calling it lambda for no particular reason)

1. All unsquaded players their lambda = their mu
2. For Squaded players:

squad_size_modifier = 1 + (player_count_in_squad * 0.05)
squad_cohesion_modifier = 1 + (historical_cohesion_of_squad * 2)
squad lambda = sum( mu_of_players_in_squad ) * squad_size_modifier * squad_cohesion_modifier

[The historical_cohesion_of_squad would be a value between 0.0 (all squad members have never played together before) and 1.0 (squad members very regularly play in the same squad with each other)]

3. The distribution of players would proceed as CCP Rattati described with the squad lambdas being placed first followed by the rest of the players without squads. All of the modifiers could be tweaked of course, this was just a rough sketch to illustrate how this could be done.

What I'm describing doesn't actually affect player's Mu score, it's a secondary stat used for balancing who is on what team. I might have a mediocre Mu score, but I'm significantly more dangerous to the opposing team if I'm squadded up, especially if I'm squadded up with people I regularly play with. This will help get more balanced fights.

Edit:
So just to clarify: The 32 players are chosen by their Mu scores. the 16 v. 16 would be organized by lambda scores.

i highly doubt this "new" approch will work for matchmakeing..

i have stated several times in several threads the only way to get balance is to..
1. create an average WP over time leaderboard system.
2. take the player average WP data
3. match players with similar WP averages with others of similar averages.
3.1 players in squads will be matched by "Squad Median" but not including empty slots in the squad, so if there is 4/6 in squad 1 it will only take the median of the 4/4 excluding the remander 2 which would mathmatically be a 0 each(unless populated by members)
4. if all turns out well.. newberrys should be matched more against newberrys while vets/pros/squads should be pit against more vets/pros/squads

The idea is very good. But I play since 3 hours and my squad (3 prima gallicus) are continuely coming into started battles and always with all red and no shield.

Every time I say "challenge guys, we return the game!", but the time we are deploying, armor is half of the max. we fight, we take objectives but it seems that all our blues are searching flowers all around the map and we are always the 3 same in the battlefront.

So it's clear, the problem is a lot of new players and they are 13 in our team. They don't know the game and what to do on the map.

So firstly, push the academy at 20 000 wp. They will have time to learn and the old players will go with new players with a few experience. Not guys who try to find the fire button on the controller.

This is basically the method we are proposing, you are adding a squad bonus modifying parameter which could be useful to tweak the system, and we already have a design for that (I just didn't mention it).

Squad size multiplier does make a difference. I know I perform consistently better in a squad of 6 than in a squad of 2. While long-term, squadding up will increase someone's average WP/minute total, from a match-by-match point of view, the matchmaker needs to detect the difference between three squads of 2 average players, and one squad of 6 average players; all in all, the larger squad will punch above their weight, not least because they'll be able to drop orbital strikes first.

Using sum instead of median makes sure that squad size is taken into account.

Correct me if I'm wrong, but if there's a bunch of players joining a battle, wouldn't it end up like so?

Step one: rank all the units

1: squad of six @ 25, so rank 150
2. squad of two @ 25, so rank 50
3. squad of two @ 25, so rank 50
4. squad of two @ 25, so rank 50
5. solo player, rank 24
6. solo player, rank 24
7-32. solo players, rank 20

If I understand correctly, the final teams would be:

Team A:
squad[6]
solo 24
solo 20 filling the roster

Team B:
squad[2]
squad[2]
squad[2]
solo 24
solo 20 filling the roster

What I suggest is that the squad of 6 is functionally more lethal than the three squads of 2, even if they contain equal-level players.

So if there were a bonus given to squads (let's say modified fibonacci: +1, +2, +3, +5, +8 pts) then the matchmaking would instead end up with the teams:

Team A:
squad[6] 150pts + 8pt squad modifier
solo 20 filling the roster

Team B:
squad[2] 50pts + 1pt squad modifier
squad[2] 50pts + 1pt squad modifier
squad[2] 50pts + 1pt squad modifier
solo 24
solo 24
solo 20 filling the roster

We actively calculate each players rank based on the result of every match, and we lovingly call that rank "Mu". Before the first battle, the player is given a rank of 25. That rank is updated after the battle, based on the weighted average of the persons duration in the battle and whether he won or lost. Simplified (the calculation is considerably more complex)

NewMu=OldMu+"Player seconds in Battle/Battle Duration" * Win/Loss

So if you lose, your Mu goes down, and if you win, your Mu goes up.

This is a modified version of elo ranking, used in many competitive sports such as chess, major league team sports and esports. Eventually your Mu will converge and stabilize around your "true skill", which is where you will win and lose equally against either players with the same Mu. However, since this is a team game, convergence will happen slower and you may find that even having the best game of your life, will not influence the match enough to secure a win. I will, however, demonstrate that it works very well to predict player skill.

The underlying problem is that after one match, everyone is very close to 25, and 2 battles, even 10 battles in, only the very best (and worst) have begun to be different from the pack.

That's why it's imperative to find a proxy for Mu for the first battles, which is what comes next in our findings.

It unclear to me if the Mu score (current and future implementation) is calculated over the lifetime of the character or per play session? Perhaps you can clarify?

I.e Will I start with Mu of 25 every time I log-in to Dust (before my first battle of the day?)

Very interested in this one.
As some that's specialized in Scouts since closed beta, I've seen each and every update (be it patch or hotfix) significantly affect the Role's performance.
As someone that's had Proficiency V in Sniper Rifles since February '13 I've only seen decline in performance from patch to patch and hotfix to hotfix.

The algorithm we are testing is

====

All squads and solo players are transformed into units

Each unit in the 32 man pool is designated with a rank, I am proposing sum of WP/s to be absolutely sure that the biggest squads with the best players are 100% not on the same side.

Top rank unit gets placed randomly on Team A or B. Team A in this case.

Units are then placed on Team B until sum of Team B exceeds sum of Team A.

Now, Units are added to Team A and switched, and so forth until all units are assigned.

We may have to increase the pool to 32+ to accommodate differing sizes of squads so we have "filler" solo players to create the 16 v 16 team. Those unfortunate enough to not get added to a team would be pushed back into the queue for the next matchmaking attempt.

====

This algorithm should always get the two best units on either team. It will also always pitch the second best unit and the third best unit against the top unit. If a solo players rank actually exceeds a squads rank, he will be placed above that squad (up to the extent that the top rank is around a 100, which equates to a squad of 6 players with a rank of 14). If top ranked players try to game the system and not join as a squad, they will still end up on opposing teams.