Jump to content


Maximally balanced MM via machine learning / AI

match making MM machine learning AI balance

  • This topic is locked This topic is locked
18 replies to this topic

Killtech #1 Posted 29 December 2019 - 12:18 AM

    Staff Sergeant

  • Player
  • 32999 battles
  • 345
  • [FLOG] FLOG
  • Member since:
    06-15-2011

As the company i work at has entering a cooperation with an university looking for research topics and new applications for machine learning. Thus we were discussing which of our common processes would be most suited to consider for such techniques. 

 

But now that i am playing WoT it crossed my mind that actually the WoT match maker would be the perfect application of these. Think about it: how many battles are fought each day? That's just an infinite pool of labeled data just sitting there to be used! These is exactly the kind of problem where AI approaches by far outclass any other solutions. And to be honest it's a extremely difficult problem for any classical algorithms and where human intuition is often more misleading then helpful. 

 

But there is also a tricky part to it: before one starts to do anything in this direction one has to define what 'balance' actually is supposed to mean and what the AI should optimize. for example one could try to let the AI find best possible team compositions such that balltes will be as far as possible from 15:0 results. so once the AI has learned enough from the training data (i.e. billions of battle results), it'll then take a list of tanks and a map as input and distribute those into two teams just in such a way that battles with similar teams resulted in the closest battles before. (allowing the AI to make decisions specific for each map will undoubtedly massively improve the MM).

 

So in this case the data labeling is trivial - the training data must just know how many tanks survived a battle in each team and the AI will always try to reproduce those battles that had the closest endings.

 

But of course with such techniques there is far more one can do. For example one could also try to make each tank have an average win rate of around 50% (The AI would be fed with all tank data, but no information about the players playing them) - regardless of how a tank is actually balanced. [EDIT: please note this is in no way a skill MM. it only averages out each tank stats individually but over the entire playerbase playing this tank, not over individual players]. This would usually mean that the AI would quickly figure out overpowered tanks from the battle data and match them together with underpowered tanks - so when playing Obj. 430U or Obj. 279 (early) one would have to carry all the IS-4s, AMX M4s and Rheimetall Panzerwagens which would end up in ones team most of the time - so the MM would actually give those powerful vehicles just the difficult tasks that they are just capable of handling - 50% of the time. So suddenly one would have to play well in those tanks to get good results. [EDIT: to make an example how that could look: WR for Obj 279 (e), for unicums it could go up to 70%, but bots would go down to even as low as 30% - so over all players that play this tank it's 50%. this is because the vehicle is so powerful that it will have huge leverage on most battles and a team will stand or fall to a large part how the player in this tank performs. for an IS-4 on the other hand unicums might only reach 55% but bots may achieve 45% - overall again 50%. as it is rare for these tanks to end up in matchups where they are the most powerful tanks in their team they will have less laverage on average thus even the best players won't be to carry with them that often. these numbers are of course made up to highlight the consequence of such an MM but I hope you get the idea]

 

and an AI could even do more: it could also take into account the current upgrade status of each vehicle and the experience level the crew (but it shouldn't know which skills they have) and balance those out. that would mean that playing a stock tank with a new crew won't cripple your team anymore since what you lack in power the MM will compensate with a proper team composition. Of course that adds a huge amount of parameters to the MM process so the AI would require a much larger data set for this to work... but given that there are billions of battles each month it's actually just a data collection problem.

 

Just when i was writing this down i got quite exited about all the possibilities. would love to work with the data that WG sits on and experiment with various approaches. ah well, anyhow, does anyone know if WG uses any kind of machine learning techniques in WoT already?


Edited by Killtech, 29 December 2019 - 04:15 AM.


1ucky #2 Posted 29 December 2019 - 12:27 AM

    Lieutenant

  • Player
  • 89606 battles
  • 1,703
  • [THRIL] THRIL
  • Member since:
    11-05-2013

Yes, someone knows that for sure.

But I wouldn't be too sure they're reading this forum.



evilchaosmonkey #3 Posted 29 December 2019 - 12:29 AM

    Captain

  • Player
  • 18383 battles
  • 2,110
  • [EIGHT] EIGHT
  • Member since:
    05-04-2013

View Post1ucky, on 28 December 2019 - 11:27 PM, said:

Yes, someone knows that for sure.

But I wouldn't be too sure they're reading this forum.

 

You may be surprised who's reading :)



Nishi_Kinuyo #4 Posted 29 December 2019 - 12:36 AM

    Lieutenant General

  • Player
  • 9945 battles
  • 6,934
  • [GUP] GUP
  • Member since:
    05-28-2011

No thanks.

If prior machine learning experiments have taught me anything, its that we should stay the **** away from it.



SoupFork #5 Posted 29 December 2019 - 12:56 AM

    Staff Sergeant

  • Player
  • 14211 battles
  • 470
  • Member since:
    11-07-2017

...and then everyone will settle around 49-51 WR and many players will leave because:

 

- The good ones realise they’ll always be carrying weak/lazy/botting players for no reward (higher WR)

- Far too many battles will end in draws or go close to 15 mins, putting people off. Imagine doing a marathon...

- The game gets stale because you’ll increasingly see the same types of enemy tanks on the same map if you grind one line.

- Botting or being inactive is “rewarded” with decent WR.



m1x_angelico #6 Posted 29 December 2019 - 12:59 AM

    Second Lieutenant

  • Player
  • 26752 battles
  • 1,448
  • [-VETO] -VETO
  • Member since:
    01-04-2015

Not sure if WG uses machine learning, I think it is best to contact them directly, especially as you are work on this as a company.

 

It is surely exciting possibility.

 

However, as you are "advocating" for averaging winrate for the sake of balance and sake quality, I recommend not going to the pinned Matchmaking Discussion Thread, because it is filled with a minority of persistent players who feel personally attacked if you mention such a possible approach. Although, it would make for an interesting discussion...

 

EDIT: It seems one was already triggered, a few seconds before I posted, so it seems they will come to you :)


Edited by m1x_angelico, 29 December 2019 - 01:00 AM.


Killtech #7 Posted 29 December 2019 - 02:04 AM

    Staff Sergeant

  • Player
  • 32999 battles
  • 345
  • [FLOG] FLOG
  • Member since:
    06-15-2011

View PostSoupFork, on 28 December 2019 - 11:56 PM, said:

...and then everyone will settle around 49-51 WR and many players will leave because:

 

- The good ones realise they’ll always be carrying weak/lazy/botting players for no reward (higher WR)

- Far too many battles will end in draws or go close to 15 mins, putting people off. Imagine doing a marathon...

- The game gets stale because you’ll increasingly see the same types of enemy tanks on the same map if you grind one line.

- Botting or being inactive is “rewarded” with decent WR.

win rate will definitively still vary a lot, but MM could make sure it's only due to skill, and not unbalanced vehicles or money (buying OP premium tanks, free-xp-ing upgrades, faster crew training). to make it absolutely clear: this is no skill MM. for as long as player information (stats, win rate, etc) is not included in the data set (and also no other data which makes it possible to derive/hint at such information) the AI won't be able to balance player skill. therefore good players will still win more frequently then bad players.

 

- good players are only good because they carry games. if they get lazy they will bleed WR very fast - especially when they play overpowered vehicles since those will have hard MM so if you play those badly you'll be able to reach abyssal WR far worse then what is possible to achieve now as an afk bot. 

- number of draws is a data labeling problem and a question how you set up your loss function - i.e. what you let the AI learn to optimize. yeah, as i stated before this is a tricky issue: one must be careful when defining what desired battles outcomes are. if you label draws a bad thing an AI will minimize those according to your weighting. so in general your concern here is very much perfectly solvable.

- stale/similar battles. this one is indeed tricky but ultimately also a question of data labeling and loss function definition. one must make sure that the AI understands that creating too similar battles is undesirable. 

- as for botting: as mentioned in my first point it will still affect WR, after all it's no skill MM. botting with OP tanks could be awarded with a 30% WR or worse but doing the same with the worst tanks in the game could actually become somewhat viable. perhaps this is a valid concern but then again making tanks somewhat balanced in the first place would negate it. 



Zylon0 #8 Posted 29 December 2019 - 02:06 AM

    Sergeant

  • Beta Tester
  • 11144 battles
  • 254
  • Member since:
    01-06-2011

There is no need for machine learning.

 

For a balanced game we need al tanks to be near the 50% winrate.

Next match tanks only against their own tier. so no longer any +- 2 or 1 tiers.

 

Just because of the fun factor we do not want player skill to be taken into acount.

Otherwise there is no reason to get better at the game.

The new Call of duty has player skill matchmaking and it sucks because of that.



azakow #9 Posted 29 December 2019 - 02:10 AM

    Major General

  • Player
  • 84701 battles
  • 5,186
  • Member since:
    05-23-2011

Anything that would reduce wait time back to less 15 sec.

Over that the last 2 years wait has increased alot.



Zylon0 #10 Posted 29 December 2019 - 02:12 AM

    Sergeant

  • Beta Tester
  • 11144 battles
  • 254
  • Member since:
    01-06-2011

View Postazakow, on 29 December 2019 - 01:10 AM, said:

Anything that would reduce wait time back to less 15 sec.

Over that the last 2 years wait has increased alot.

People do not want to play with thier tier 8 vs tier x all the time.

That was the whole reason the MM changed recently.

 

Of course it was faster MM, but that is not a priority over a fair and fun game.


Edited by Zylon0, 29 December 2019 - 02:12 AM.


mtnm #11 Posted 29 December 2019 - 08:03 AM

    Corporal

  • Player
  • 3848 battles
  • 138
  • Member since:
    11-30-2016
Very nice reading. I am 120% sure WG would use machine learning only to maximize their profits.

gunslingerXXX #12 Posted 29 December 2019 - 08:41 AM

    Colonel

  • Player
  • 13577 battles
  • 3,664
  • [GUNSL] GUNSL
  • Member since:
    11-16-2014

Interesting idea. I'm not sure on how it will work for the game experience though (or WG profit for that matter). Part of the fun in the game is getting better tanks and crew, this is not really desirable anymore then. Basically you need to find tanks in which you perform above average to do well. 

Also to have more 15 minutes battles might become boring. I'm interested to see the ideas WG wants to implement to decrease the nunber of landslide battles.



jabster #13 Posted 29 December 2019 - 10:51 AM

    Field Marshal

  • Beta Tester
  • 12790 battles
  • 26,546
  • [WSAT] WSAT
  • Member since:
    12-30-2010

Technically I do believe you could balance, within reason, all those variables that aren’t player skill. What I’m not convinced of is that it will achieve the results that you are hoping for. There’s been some analysis of looking at the relationship between (un)balanced teams (in the context of skill) and (un)balanced results. This points to the correlation between them is at best very weak. So if balancing teams, by skill, has almost no effect I don’t see balancing maps, tanks etc. would either.

 

My personal opinion is that the overriding factor is the game format combined with how inconsistent the average player is in an individual battle. That isn’t something you can balance.



evilchaosmonkey #14 Posted 29 December 2019 - 11:03 AM

    Captain

  • Player
  • 18383 battles
  • 2,110
  • [EIGHT] EIGHT
  • Member since:
    05-04-2013

View Postjabster, on 29 December 2019 - 09:51 AM, said:

My personal opinion is that the overriding factor is the game format combined with how inconsistent the average player is in an individual battle. That isn’t something you can balance.

 

Exactly this.  The model will narrow the margin down a fair bit, and will give a good opportunity for a balanced game.  The model can never predict the variables that happen within a game.

Building the model isn't hard, nor is getting it to predict an outcome to around 70% ABS accuracy - done already, well at least for an individual player and how they would perform in a given setup before game start.

 

Bigger question is, if you could predict an outcome to 90%++ before the match would you want it?  Do people want games that are almost predetermined?  Doubt it.



Wintermute_1 #15 Posted 29 December 2019 - 11:08 AM

    Captain

  • Player
  • 52464 battles
  • 2,448
  • Member since:
    11-25-2013
If AI could learn to review 'reports' based on correctly assessed reports by WG staff it would be more helpful. Instantly handing out sanctions for spurious reporters and to those who actually break in game rules. The report function might actually be worth something if it wasn't ignored because it involves an unrealistic workload for human employees.

Scanmen #16 Posted 29 December 2019 - 11:14 AM

    Sergeant

  • Player
  • 25981 battles
  • 278
  • [WNB] WNB
  • Member since:
    10-14-2011

 

.


Edited by Scanmen, 29 December 2019 - 11:17 AM.


jabster #17 Posted 29 December 2019 - 11:32 AM

    Field Marshal

  • Beta Tester
  • 12790 battles
  • 26,546
  • [WSAT] WSAT
  • Member since:
    12-30-2010

View Postevilchaosmonkey, on 29 December 2019 - 10:03 AM, said:

 

Exactly this.  The model will narrow the margin down a fair bit, and will give a good opportunity for a balanced game.  The model can never predict the variables that happen within a game.

Building the model isn't hard, nor is getting it to predict an outcome to around 70% ABS accuracy - done already, well at least for an individual player and how they would perform in a given setup before game start.

 

Bigger question is, if you could predict an outcome to 90%++ before the match would you want it?  Do people want games that are almost predetermined?  Doubt it.


For the first part, I’m not sure I agree. In my opinion in the long run, for the average player overall performance and actual skill have a good correlation but I don’t feel that’s true at the individual battle level. To put it simply the main difference between me dying in the first two minutes and getting a top gun is far more to do with luck than my judgement.



evilchaosmonkey #18 Posted 29 December 2019 - 11:49 AM

    Captain

  • Player
  • 18383 battles
  • 2,110
  • [EIGHT] EIGHT
  • Member since:
    05-04-2013

View Postjabster, on 29 December 2019 - 10:32 AM, said:


For the first part, I’m not sure I agree. In my opinion in the long run, for the average player overall performance and actual skill have a good correlation but I don’t feel that’s true at the individual battle level. To put it simply the main difference between me dying in the first two minutes and getting a top gun is far more to do with luck than my judgement.

 

I hoped that's more or less what I said.  Perhaps not clear enough.  AI (at least neural networks) are surprisingly good at being able to predict a battle outcome, at least I thought 70% ABS accuracy was good.  The 30% is what happens in a battle.



Homer_J #19 Posted 29 December 2019 - 12:22 PM

    Field Marshal

  • Moderator
  • 33157 battles
  • 36,664
  • [WJDE] WJDE
  • Member since:
    09-03-2010

View Postm1x_angelico, on 28 December 2019 - 11:59 PM, said:

 

However, as you are "advocating" for averaging winrate for the sake of balance and sake quality, I recommend not going to the pinned Matchmaking Discussion Thread,

 

I considered merging the threads but that would have OP confused and lonely looking for where it went so I will settle for closing this one.  Feel free to continue the discussion in the pinned thread a suggested above.






1 user(s) are reading this topic

0 members, 1 guests, 0 anonymous users