Guild Wars 2

DISCLAIMER I know almost nothing about machine learning, and I am writing this right after I woke up (Gotta strike the iron while its hot!). I did study CS in school (But once I entered the workforce, I realized how little I knew!).

Inspired by every time I see a post on GW2 or League saying: “how is this matchmaking fair!” Are there times where matchmaking is clearly unfair that a human would be able to tell? If so can we use machine learning to be the human that prevents those matchups?

So I thought to myself. Glicko / MMR / Elo, are all systems that really are designed to pair 1 person to 1 other person. They work well in that instance. However, since all those systems are based on one or two metrics (“win-rate” and time), they aren’t really good for pairing teams vs teams. There are various aspects that are lost when you only pair on MMR. Game acquired metrics such as: how balanced a game is, how long a game is, how long it took to queue a game together, how many negative / positive words were said in the game (also consider how many shotcalling / team communication words were said) are lost. User-feedback acquired perspectives of how balanced the game was, how fun the game was, how toxic / friendly their teammates were are lost. Per-user acquired metrics of afk rate, and word-usage are also lost. Would it be possible to improve matchmaking, while still keeping matchmaking inherently fair, by including some or all of the above metrics?

I believe that machine learning might be able to capture some or all of the above metrics, perhaps by including them as individual weights for each player. Each team v. team paring can be judged after the fact based on randomly selected user feedback and by analyzing the game. I think this could lead to games that are funner and more balanced.

However there are some concerns. Primarily, if such a system could really be fair. Clearly, MMR could be used to create a pool of players that machine learning is then used to select from. But weather or not this is as fair as the (I assume) random selection before is hard to say. Players may not like being paired in a manner that they cannot clearly tell is fair.

Other applications

General automated team creation. Imagine creating a development team from a pool of crowdsourced / freelance developers (Hate contest-based crowdsourcing. But that’s another topic). Could create an HR world where you hire based on attempting to achieve a “team” that based on machine learning should be optimal. Rather than hiring random individuals. Or, given a set of individuals, could try and find the remaining team that would work the best with those individuals. Of course initially such a system might only be used only as a suggestion on top of analyzing all possible candidates. But, if it proves to be effective, later on it could be phased in as a primary method of acquisition.

*Side thought: everywhere I say machine learning. Might be appropriate to replace with neural networks.

Midi.8359:

So I thought to myself. Glicko / MMR / Elo, are all systems that really are designed to pair 1 person to 1 other person. They work well in that instance. However, since all those systems are based on one or two metrics (“win-rate” and time), they aren’t really good for pairing teams vs teams.

Many people say this. But it’s all based on feelings, not any hard evidence.

Matchmaking is working pretty well, but suffers from three things:

1. Low population at off-hours, especially for higher skill levels. The matchmaker will give players a poor match over no match at all.
2. Players are unable to gauge their own skill level. Many think they’re better than they actually are.
3. HoT builds have a lot of “first order optimal” design which cause players to hit a skill wall. When something a player has done forever no longer works, the typical response is to blame other factors like matchmaking.

You do realize that machine learning is still a highly experimental field? Developing a machine learning matchmaker would require both a research team and many years.

Also you should probably do some research of your own. Glicko is hardly new to the world of team games, and there has been much research done on the effectiveness of various matchmaker algos in team games. Honestly my own research tends to indicate that ArenaNet’s setup is sound, however it’s plausible that there could be bugs.

not needed complex machine learning algorism and other kittens:
the solution is easy:
-class mmr
-balanced classes or knowing the imbalance such if team A have X team B must have X or if not posible Y
in this way matches only will be “broken” when some player/s brings to the table some creative and funcional “out off the box” build

Exedore.6320:

Many people say this. But it’s all based on feelings, not any hard evidence.

Hrm… I should have added that for me personally. I find the matchmaking fine in terms of balance. Can lose to where I’m better than most people I play with. Win to where I’m worse. However, that doesn’t mean it couldn’t be better for others and in aspects other than balance. Such as queue time.

Crinn.7864:

You do realize that machine learning is still a highly experimental field? Developing a machine learning matchmaker would require both a research team and many years.

Yup and yup.

Crinn.7864:

Also you should probably do some research of your own. Glicko is hardly new to the world of team games, and there has been much research done on the effectiveness of various matchmaker algos in team games. Honestly my own research tends to indicate that ArenaNet’s setup is sound, however it’s plausible that there could be bugs.

Hrm… I looked at the public Glicko algorithm for GW2. Like I said above, personally I find it matchmaking rather fine. However from looking at it’s input parameters it doesn’t consider some of the aspects mentioned above. Also both in League and GW2 I hear about how when you start to get to the top of the ladder things start to break down. Would be interested if there was a way to fix that.

Tbh the pvp system isnt bad the issue is that it works when the population is big enough. The issues are that the incentives arent good enough to draw ppls attention and that theres not really a thing as a big event to hype ppl and make them look forward to. Also mismanagement from anets part.

Crinn.7864:

Honestly my own research tends to indicate that ArenaNet’s setup is sound, however it’s plausible that there could be bugs.

I would mostly agree.
The only shortcoming I’ve seen is how it expands rating ranges: it keeps it static for several minutes and then grows quickly. A more gradual growth would be more appropriate.

Elo – How many points do you award for a win? This “k” value has no easy answer and has been a problem in chess.

Glicko – What is the player’s variance and how does it change per game. These are also kind of guesswork.

A problem with machine learning for match making is the feedback. Lets say glicko gives red team a 30% chance to win. Assume red does win. How do we read into that? Blue could win the next two, but those matches will never be played.

It’s difficult and sometimes it seems that the best solutions are kind of weird.

Sc2 uses game metrics to aid matchmaking. Note, these graphs were made by the community.

https://m.imgur.com/GFJk3
https://m.imgur.com/jmMNR

I read your post twice to try and understand, but I must admit that I propably did not, at least not fully. So I will just plain and simply ask, feel free to correct me: What you suggest in your post is that the adjusted match maker takes, by a great deal, in to account a very personality/character of each possible candidate for a match? If so, wouldn’t that create nearly infinite ques that would in the end, end up being manipulated? Also, wouldn’t it totally exclude certain individuals from the matching because they would fit to no one?