I think the OP has got his answer, there is what? 4-5 other threads that go into more Detail about how Hive really works. The simple answer - It dose the job we need it to do. Some say not good enough, some say it dont work at all, but after over a year off being live, its doing its job, and "some" games that force random are more even than before.
I barely play one good game in months. The last one was 2 months ago on hellarious. Any game with FET was a disaster for both teams precisely because of improperly balanced team. Imbalanced in a way that always lead to stacking or quit/f4. Yes it works doing the FET. Problem is the numbers used aren't representative of the player skill.
@moultano : I am still waiting for the publication of your graphs (or charts); with proper documentation on methods and code. Something that can be reproduced. So far no one seen it. It's not only that people have to prove it doesn't work but for the submitting group to demonstrate it does. This something should depicts the ns2 player career in an environment similar to what we have.
About god. So it's 1 person with a godlike aim... funny that aim is not taken under consideration as a skill. That godlike aim person is about 0.5% of the remaining players (<200). Quite significant indeed. Still... the average people; will have a number of win loose that is 40 to 60% w/l. Good luck guys for trying to climb Olympus mons.
Stackers could have been contained if this thing didn't brought farmers in the process.
"Gee I bet this thread is going to involve 1 post explaining the skill system in simple terms, and then several pages of certain people showing that they do not understand how the skill system works at all."
Because its not a "skill" as persay its the likely hood off said team winning, what also dose not help is when the vote happens, players can F4 right away, and switch teams.
Somthing that may be usful is that when a vote is happening - even if in options a player can choose what side they would prefer to play on, then look at what team they played on last, if they have just played 3 rounds in a row as aliens then they should have a better chance off playing marines, One thing i hate is when i get randomed to a team after playing 4-5 rounds on the same team, HATE IT.
But yea, if players could choose what team they prefer, then maybe less F4ing?
Prob Not tho.
Not sure if the 'shuffle' function in servers is ns2 coded or a server thing...but I think the hive score/shuffle system would work a lot better if it took rookies into account. I notice it's still difficult to have even games because the shuffle puts the really high hive scores with the really low hive scores, but doesn't separate the rookies which I think is essential in this game. Rookies a lot of the time are put all on one team or 75% or more of the rookies are on one team. I understand why it happens, but it doesn't work and is a glaring issue in the hive score/shuffle system. It leads to bad games, and it leads to people not liking the shuffle system. It also can't be fun for new players when you're on a team that is getting destroyed.
I would like to see the shuffle system put an even number of rookies on each team(close as possible) and then separate the rest of the players/teams via hive score. I think if it could do both of those things together then we could have better/even pub games. Thoughts?
@Deck_ I think you might be on to something. I know this might come off as elitist - but you could "force" anyone under a certain skill number to have "0" skill factor. Then distribute those evenly, and then distribute everyone else normally.
It is often times the top 25% (if not less) that carry a team.
It might also be worth a try grouping people into tiers based on their hive score and balance by that rather than average value and see how that works out.
@Deck_ I think you might be on to something. I know this might come off as elitist - but you could "force" anyone under a certain skill number to have "0" skill factor. Then distribute those evenly, and then distribute everyone else normally.
It is often times the top 25% (if not less) that carry a team.
Yes, and it is also the bottom part of a team that can ruin a game. I really like that idea. I'm not sure if you need to force them to have a certain skill. You just need the shuffle system to take into account who is a rookie and who is not. They have to be treated different so they are spread evenly between both teams. I just want to see this implemented into shuffle, I think it has a much better chance of working. Shuffle system is great in theory, but the rookie factor is the missing link.
It might also be worth a try grouping people into tiers based on their hive score and balance by that rather than average value and see how that works out.
this would also work, but not sure which is easier to implement. This form of shuffle or the separate rookies first before the rest of the players shuffle. I think both would be an improvement.
Person8880Join Date: 2013-01-02Member: 177167Members, Squad Five Blue
I'd like to clear up a few things I read often about team shuffling.
Firstly, there's two kinds you've probably seen. The game's "Force Even Teams" vote, and the "Shuffle" vote. The "Shuffle" vote is from Shine's team balancing plugin (which I wrote), and it uses a significantly different method from the game's.
Force even teams attempts to come up with teams that have the least difference in total skill between them. This can lead to flawed results when teams are not the same size.
Shine's shuffling attempts one or two things, depending on configuration:
Minimise the difference in average skill between the two teams. The main reason for average instead of sum is that average accounts for when one team has one more player than the other. Using the sum places more higher skill players clumped on the lower count team to try and make up for the difference, the average is unaffected by this.
If configured (which it is by default), also attempt to minimise the difference in standard deviation between the two teams. This will account for when two teams have similar average but one is lots of high and low skill while the other is all medium skill. Up until today there was a small oversight in this that would sometimes leave the teams worse than they should be, but that should be corrected now.
I've also seen people mention locking teams, Shine's shuffle vote can be configured to do this. However, people will often leave and rejoin to try and unbalance the teams. The plugin does attempt to stop this by placing players onto the lower skill team if it can, but it can't always do this. This is more an attitude problem and is not something easily fixed automatically without seriously annoying players.
Another thing I've seen people mention is that the skills are not linear scaling. This may be so, but it doesn't affect team balancing at all. The balancing is done based on relative rankings comparing players against each other. No matter what (reasonable) transformation you apply, as long as it's the same to all players the relative rankings remain unchanged. You could remove data by rounding all skills off to a given precision (say, the nearest 500 points), which would perhaps change the sorting slightly, but it would still place high skill players as high skill, and low skill as low skill. The only possible difference would be the distribution of those near each other, which often matters little anyway as it's the extremes that cause the most issues.
At the end of the day, you can play with the numbers as much as you want, people will never be happy with it. There's only so much data you can contain in a single number, and a human is a whole lot more complex than that.
[*]Force even teams attempts to come up with teams that have the least difference in total skill between them. This can lead to flawed results when teams are not the same size.
[*]Shine's shuffling attempts one or two things, depending on configuration:
Minimise the difference in average skill between the two teams. The main reason for average instead of sum is that average accounts for when one team has one more player than the other. Using the sum places more higher skill players clumped on the lower count team to try and make up for the difference, the average is unaffected by this.
Yeah, we should really do this Force Even Teams too, this is an oversight.
[*]If configured (which it is by default), also attempt to minimise the difference in standard deviation between the two teams. This will account for when two teams have similar average but one is lots of high and low skill while the other is all medium skill. Up until today there was a small oversight in this that would sometimes leave the teams worse than they should be, but that should be corrected now.
Tell me more! What function is it optimizing exactly? I wanted to put something like this in Force Even Teams as well, but I haven't had time to sit down and work on it properly.
I was thinking of minimizing the sum of variances, but got stuck thinking that this is in most cases the exact opposite goal of balancing the teams, but minimizing the difference of variances seems like it could work much better.
Edit: I just had a thought. This is total overkill, but this could be a good theoretical framework. The goal of making the teams even is to make the teams as indistinguishable as possible. This concept, of making things indistinguishable, actually has a precise definition from information theory: Mutual information. So treat it as two random variables , the team selection in [0,1], and the skill values, which you can model as gaussians. The goal is to minimize the mutual information between the team selection variable, and the skill distribution variables. This metric should combine the difference of means and the difference of variances into one score to optimize for! I actually had to work this out at work for another application, but this seems like a perfect context for it.
Another thing I've seen people mention is that the skills are not linear scaling. This may be so, but it doesn't affect team balancing at all. The balancing is done based on relative rankings comparing players against each other. No matter what (reasonable) transformation you apply, as long as it's the same to all players the relative rankings remain unchanged. You could remove data by rounding all skills off to a given precision (say, the nearest 500 points), which would perhaps change the sorting slightly, but it would still place high skill players as high skill, and low skill as low skill. The only possible difference would be the distribution of those near each other, which often matters little anyway as it's the extremes that cause the most issues.
Can you talk a little more about what the algorithm does? The hive model is designed for skills to be added, which might be a bad model, but the scores are optimized assuming that's true, so adding them up so that the average is the same should be the right thing to do.
I regularly check other peoples hive scores on my team to get an "idea" of how well we are going to perform, specially if I command, if I cannot see someones hive score I have no idea how well our team will potentially play out and adjust accordingly.
My mind is also at ease if I know a 1500+ commander jumps into the chair than a player who has only played 20 hours.
I actually blame this approach on the tons of games we have where people prematurely concede. It's a self-fulfilling prophecy. People look at the hive scores and conclude that they can't win and so they fuck around waiting for the round to end instead of playing to win.
i don't know, oftentimes, you know the game is over. And it's really really really fucking over.
you're on 1 res, they've got every fucking lane covered by at least one guy.
you've been trying to sneak a skulk out for the better part of 5 minutes. They've been poking in, and you're pretty sure the other guys would notice an open tunnel/lack of defense.
and the guy in the other chair is almost certainly on point with his beacons because he's been on point with his scans so far.
and you can't get 2 skulks to kill a guy on an upgrade let alone into a fucking tunnel.
Yeah, when you're been outplayed, like really outplayed, there's a 5 percent chance of a comeback, and a 95 percent chance of dragging it out for another 10-15 minutes. just to lose while egg locked.
apologies, included another line somewhere for no reason
Person8880Join Date: 2013-01-02Member: 177167Members, Squad Five Blue
It's just looking at the current average (and standard deviation) of both teams, and checking whether swapping each possible pair of players reduces the difference in average skill (and the difference in standard deviation if possible). By difference, I mean literally |Team1Average - Team2Average| and similarly for standard deviation.
The best swap it finds that reduces the difference in team averages, and if possible, reduces the average in standard deviations (though it will allow the standard deviation to rise slightly if the difference in averages can be improved) is the one that is performed. It keeps at this until it can't find a swap that reduces the difference in average skill between the two teams anymore.
I can't say that it's the best thing ever, but from testing it produces consistent and (to the naked eye) correct enough teams. It's been a while since I did any formal statistics though so I'm sure that there's further improvements possible.
@amoral - Hey I'm not above conceding. But some people act defeated before the game even begins because of differences in hive skill. < That's what I'm talking about.
However, I will add that I have had games where we have the entire map as marines and for what ever reason we lose nanogrid on veil, and everybody wants to insta-quit. In spite of the fact that we have 3/3 jetpacks and 4+ rts. So the average pubber is not as good at judging when a game is over as they like to think in my experience.
The best swap it finds that reduces the difference in team averages, and if possible, reduces the average in standard deviations (though it will allow the standard deviation to rise slightly if the difference in averages can be improved) is the one that is performed.
Yeah, this is actually the detail I was curious about. How much difference in average skill is it willing to trade off for an improvement to the difference in standard deviation?
The idea is based on what you are saying about making the mean and standard deviations as close as possible. It just defines that "as close as possible" a little more formally. The idea is that you model the skill distribution of the overall server as a normal distribution. Then you model the skill distribution of each team as a normal distribution. The goal in balancing the teams is to make the sum of KL Divergences between each team's skill distribution and the server's skill distribution as low as possible. That is, each team should have a skill distribution roughly the same as the overall distribution on the server.
The best swap it finds that reduces the difference in team averages, and if possible, reduces the average in standard deviations (though it will allow the standard deviation to rise slightly if the difference in averages can be improved) is the one that is performed.
Yeah, this is actually the detail I was curious about. How much difference in average skill is it willing to trade off for an improvement to the difference in standard deviation?
It's the other way round. It will never ignore a swap that lowers the difference in average skill, but it will allow a swap that might increase the difference in standard deviation by a small bit if it still reduces the difference in averages. The reason is it often got stuck trying to reduce standard deviation and gave up when teams were nowhere near balanced. While it could reduce the difference in averages, the standard deviation difference couldn't get any better.
"Gee I bet this thread is going to involve 1 post explaining the skill system in simple terms, and then several pages of certain people showing that they do not understand how the skill system works at all."
I wasn't disappointed.
I actually got my question answered quite well by the second person to post on this thread. So I don't understand what everyone else is babbling about. Something about how a better system is needed and that Hive Skill rating should be hidden. I think it works mighty fine right now.
The best swap it finds that reduces the difference in team averages, and if possible, reduces the average in standard deviations (though it will allow the standard deviation to rise slightly if the difference in averages can be improved) is the one that is performed.
Yeah, this is actually the detail I was curious about. How much difference in average skill is it willing to trade off for an improvement to the difference in standard deviation?
It's the other way round. It will never ignore a swap that lowers the difference in average skill, but it will allow a swap that might increase the difference in standard deviation by a small bit if it still reduces the difference in averages. The reason is it often got stuck trying to reduce standard deviation and gave up when teams were nowhere near balanced. While it could reduce the difference in averages, the standard deviation difference couldn't get any better.
Is it possible to have the system separate the rookies evenly before doing the standard shuffling that you described? I would like to see if it works better. Or let's say separate the people evenly with a hive score below a certain point and then do your average thing with the rest of the players.
Perhaps you remember the previous system that did exactly what you are advocating for? It rewarded based on points in the round, and resulted in spawn camping for hours and emptying servers.
A high round point is related to effective effort during gameplay. Players who can imprint effort and are successful in doing so (a metric already measured by the points) tend to be helpful to the team either by killing RTs/structures, by killing high life forms, by killing many low life forms which all eventually translates into a win.
This view captures rookies / avg players / good players on pub matches. I believe it can provide better rank discernment (improving all the benefits it brings) and also address the "categorization" issue.
Could the deviation from the avg team round points be used to weight (a non linear coefficient can avoid farming) the hive score points won/lost? Can it be tested?
The best swap it finds that reduces the difference in team averages, and if possible, reduces the average in standard deviations (though it will allow the standard deviation to rise slightly if the difference in averages can be improved) is the one that is performed.
Yeah, this is actually the detail I was curious about. How much difference in average skill is it willing to trade off for an improvement to the difference in standard deviation?
It's the other way round. It will never ignore a swap that lowers the difference in average skill, but it will allow a swap that might increase the difference in standard deviation by a small bit if it still reduces the difference in averages. The reason is it often got stuck trying to reduce standard deviation and gave up when teams were nowhere near balanced. While it could reduce the difference in averages, the standard deviation difference couldn't get any better.
Is it possible to have the system separate the rookies evenly before doing the standard shuffling that you described? I would like to see if it works better. Or let's say separate the people evenly with a hive score below a certain point and then do your average thing with the rest of the players.
Shine's hive score based shuffle actually already does this. Players with a hive score of 0 get shuffled using the fall-back method (by default last rounds k/a/d).
Perhaps you remember the previous system that did exactly what you are advocating for? It rewarded based on points in the round, and resulted in spawn camping for hours and emptying servers.
A high round point is related to effective effort during gameplay. Players who can imprint effort and are successful in doing so (a metric already measured by the points) tend to be helpful to the team either by killing RTs/structures, by killing high life forms, by killing many low life forms which all eventually translates into a win.
This view captures rookies / avg players / good players on pub matches. I believe it can provide better rank discernment (improving all the benefits it brings) and also address the "categorization" issue.
Could the deviation from the avg team round points be used to weight (a non linear coefficient can avoid farming) the hive score points won/lost? Can it be tested?
The fuck do you care for, you just leave the game and do whatever it is to make the hive score loss go away.
Perhaps you remember the previous system that did exactly what you are advocating for? It rewarded based on points in the round, and resulted in spawn camping for hours and emptying servers.
A high round point is related to effective effort during gameplay. Players who can imprint effort and are successful in doing so (a metric already measured by the points) tend to be helpful to the team either by killing RTs/structures, by killing high life forms, by killing many low life forms which all eventually translates into a win.
This view captures rookies / avg players / good players on pub matches. I believe it can provide better rank discernment (improving all the benefits it brings) and also address the "categorization" issue.
Could the deviation from the avg team round points be used to weight (a non linear coefficient can avoid farming) the hive score points won/lost? Can it be tested?
The fuck do you care for, you just leave the game and do whatever it is to make the hive score loss go away.
wooo, we got some salt there.
you're getting awfully close to a personal attack there.
The best swap it finds that reduces the difference in team averages, and if possible, reduces the average in standard deviations (though it will allow the standard deviation to rise slightly if the difference in averages can be improved) is the one that is performed.
Yeah, this is actually the detail I was curious about. How much difference in average skill is it willing to trade off for an improvement to the difference in standard deviation?
It's the other way round. It will never ignore a swap that lowers the difference in average skill, but it will allow a swap that might increase the difference in standard deviation by a small bit if it still reduces the difference in averages. The reason is it often got stuck trying to reduce standard deviation and gave up when teams were nowhere near balanced. While it could reduce the difference in averages, the standard deviation difference couldn't get any better.
Is it possible to have the system separate the rookies evenly before doing the standard shuffling that you described? I would like to see if it works better. Or let's say separate the people evenly with a hive score below a certain point and then do your average thing with the rest of the players.
Shine's hive score based shuffle actually already does this. Players with a hive score of 0 get shuffled using the fall-back method (by default last rounds k/a/d).
Yes, the players with hive score of 0 get shuffled, but do they get spread evenly between both teams? I have noticed that all rookies end up on one team or most of the rookies are on one team after a shuffle (I think it was shuffle and not the other one). Are you saying they are spread evenly between both teams or they only factor into the average shuffle?
Yes, the players with hive score of 0 get shuffled, but do they get spread evenly between both teams? I have noticed that all rookies end up on one team or most of the rookies are on one team after a shuffle (I think it was shuffle and not the other one). Are you saying they are spread evenly between both teams or they only factor into the average shuffle?
They'll always be spread evenly, as they're allocated after the players with meaningful skill values have been allocated out. These 0 skill value players fill up the remaining slots loosely based on which team they'll benefit more by past KDR/score or pure random depending on how it's configured.
I think the OP has got his answer, there is what? 4-5 other threads that go into more Detail about how Hive really works. The simple answer - It dose the job we need it to do. Some say not good enough, some say it dont work at all, but after over a year off being live, its doing its job, and "some" games that force random are more even than before.
I barely play one good game in months. The last one was 2 months ago on hellarious. Any game with FET was a disaster for both teams precisely because of improperly balanced team. Imbalanced in a way that always lead to stacking or quit/f4. Yes it works doing the FET. Problem is the numbers used aren't representative of the player skill.
@moultano : I am still waiting for the publication of your graphs (or charts); with proper documentation on methods and code. Something that can be reproduced. So far no one seen it. It's not only that people have to prove it doesn't work but for the submitting group to demonstrate it does. This something should depicts the ns2 player career in an environment similar to what we have.
About god. So it's 1 person with a godlike aim... funny that aim is not taken under consideration as a skill. That godlike aim person is about 0.5% of the remaining players (<200). Quite significant indeed. Still... the average people; will have a number of win loose that is 40 to 60% w/l. Good luck guys for trying to climb Olympus mons.
Stackers could have been contained if this thing didn't brought farmers in the process.
Last check; 194.
The skill is pretty well indicitive of who will win, upsets don't happen much, except in the last month or two where setting NEW rookies to 1000 while leaving old rookies at 0 completely FUBAR'ed the existing ratings. That rating change effectively meant that more skilled, but generally inexperienced players were significantly lower in hive rating than completely oblivious rookies, and your typical strong pub player has a rating so close to a fresh rookie (~1500 compared to 1000) that it doesn't make a significant impact on force even. The ratings might be able to even out eventually, but it'll be a long time going and every time a low rated "average" player comes back to play the game it will delay the equilibrium even more.
Anyway, outside of the RECENT changes, the problem with hive skill and force-even is that even if it's a close rating and "anyone's game", regardless of which side actually wins the end result is usually still a stomp. I don't think those games are actually "stacked", I think it's a product of rapid snowballing paired with pub players not having the know-how to turn the tables once it gets going. One team wins the first couple engagements by shear coincidence and secures enough of a lead that a pub team of of equal skill simply can't hope to get a leg up on them from that point on.
I think the OP has got his answer, there is what? 4-5 other threads that go into more Detail about how Hive really works. The simple answer - It dose the job we need it to do. Some say not good enough, some say it dont work at all, but after over a year off being live, its doing its job, and "some" games that force random are more even than before.
I barely play one good game in months. The last one was 2 months ago on hellarious. Any game with FET was a disaster for both teams precisely because of improperly balanced team. Imbalanced in a way that always lead to stacking or quit/f4. Yes it works doing the FET. Problem is the numbers used aren't representative of the player skill.
@moultano : I am still waiting for the publication of your graphs (or charts); with proper documentation on methods and code. Something that can be reproduced. So far no one seen it. It's not only that people have to prove it doesn't work but for the submitting group to demonstrate it does. This something should depicts the ns2 player career in an environment similar to what we have.
About god. So it's 1 person with a godlike aim... funny that aim is not taken under consideration as a skill. That godlike aim person is about 0.5% of the remaining players (<200). Quite significant indeed. Still... the average people; will have a number of win loose that is 40 to 60% w/l. Good luck guys for trying to climb Olympus mons.
Stackers could have been contained if this thing didn't brought farmers in the process.
Last check; 194.
The skill is pretty well indicitive of who will win, upsets don't happen much, except in the last month or two where setting NEW rookies to 1000 while leaving old rookies at 0 completely FUBAR'ed the existing ratings. That rating change effectively meant that more skilled, but generally inexperienced players were significantly lower in hive rating than completely oblivious rookies, and your typical strong pub player has a rating so close to a fresh rookie (~1500 compared to 1000) that it doesn't make a significant impact on force even. The ratings might be able to even out eventually, but it'll be a long time going and every time a low rated "average" player comes back to play the game it will delay the equilibrium even more.
Anyway, outside of the RECENT changes, the problem with hive skill and force-even is that even if it's a close rating and "anyone's game", regardless of which side actually wins the end result is usually still a stomp. I don't think those games are actually "stacked", I think it's a product of rapid snowballing paired with pub players not having the know-how to turn the tables once it gets going. One team wins the first couple engagements by shear coincidence and secures enough of a lead that a pub team of of equal skill simply can't hope to get a leg up on them from that point on.
Didn't it start out that way? I know when I alt accounted it I needed to climb from 0 while some rookies were at 1000, that was around half a year ago?
Yes, the players with hive score of 0 get shuffled, but do they get spread evenly between both teams? I have noticed that all rookies end up on one team or most of the rookies are on one team after a shuffle (I think it was shuffle and not the other one). Are you saying they are spread evenly between both teams or they only factor into the average shuffle?
They'll always be spread evenly, as they're allocated after the players with meaningful skill values have been allocated out. These 0 skill value players fill up the remaining slots loosely based on which team they'll benefit more by past KDR/score or pure random depending on how it's configured.
Is it zero or close to zero? Rookies can luck into a win just by joining the correct team and won't always be at zero.
Yes, the players with hive score of 0 get shuffled, but do they get spread evenly between both teams? I have noticed that all rookies end up on one team or most of the rookies are on one team after a shuffle (I think it was shuffle and not the other one). Are you saying they are spread evenly between both teams or they only factor into the average shuffle?
They'll always be spread evenly, as they're allocated after the players with meaningful skill values have been allocated out. These 0 skill value players fill up the remaining slots loosely based on which team they'll benefit more by past KDR/score or pure random depending on how it's configured.
The skill is pretty well indicitive of who will win, upsets don't happen much, except in the last month or two where setting NEW rookies to 1000 while leaving old rookies at 0 completely FUBAR'ed the existing ratings.
New rookies don't get 1000 skill. The only "rookies" getting 1000 skill are people that already played before the hive reset that was like one year ago. So if an old player comes back, he will have 1000 skill. You can argue with that number, but the idea was that if somebody owns ns2 since several years, he is likely not a rookie with 0 skill and thus is started at 1000.
If you see any evidence that completely new rookies do indeed have 1000 skill, please link that users account. That would be a bug and is not intended.
Comments
I barely play one good game in months. The last one was 2 months ago on hellarious. Any game with FET was a disaster for both teams precisely because of improperly balanced team. Imbalanced in a way that always lead to stacking or quit/f4. Yes it works doing the FET. Problem is the numbers used aren't representative of the player skill.
@moultano : I am still waiting for the publication of your graphs (or charts); with proper documentation on methods and code. Something that can be reproduced. So far no one seen it. It's not only that people have to prove it doesn't work but for the submitting group to demonstrate it does. This something should depicts the ns2 player career in an environment similar to what we have.
About god. So it's 1 person with a godlike aim... funny that aim is not taken under consideration as a skill. That godlike aim person is about 0.5% of the remaining players (<200). Quite significant indeed. Still... the average people; will have a number of win loose that is 40 to 60% w/l. Good luck guys for trying to climb Olympus mons.
Stackers could have been contained if this thing didn't brought farmers in the process.
Last check; 194.
"Gee I bet this thread is going to involve 1 post explaining the skill system in simple terms, and then several pages of certain people showing that they do not understand how the skill system works at all."
I wasn't disappointed.
Somthing that may be usful is that when a vote is happening - even if in options a player can choose what side they would prefer to play on, then look at what team they played on last, if they have just played 3 rounds in a row as aliens then they should have a better chance off playing marines, One thing i hate is when i get randomed to a team after playing 4-5 rounds on the same team, HATE IT.
But yea, if players could choose what team they prefer, then maybe less F4ing?
Prob Not tho.
You can do that, just join a team before the vote ends. The game will try to keep you in them team if possible.
Ok, thats great, just pointing out that it only happens about 10% off the time, i get switched most off the time.
I would like to see the shuffle system put an even number of rookies on each team(close as possible) and then separate the rest of the players/teams via hive score. I think if it could do both of those things together then we could have better/even pub games. Thoughts?
It is often times the top 25% (if not less) that carry a team.
Yes, and it is also the bottom part of a team that can ruin a game. I really like that idea. I'm not sure if you need to force them to have a certain skill. You just need the shuffle system to take into account who is a rookie and who is not. They have to be treated different so they are spread evenly between both teams. I just want to see this implemented into shuffle, I think it has a much better chance of working. Shuffle system is great in theory, but the rookie factor is the missing link.
this would also work, but not sure which is easier to implement. This form of shuffle or the separate rookies first before the rest of the players shuffle. I think both would be an improvement.
Firstly, there's two kinds you've probably seen. The game's "Force Even Teams" vote, and the "Shuffle" vote. The "Shuffle" vote is from Shine's team balancing plugin (which I wrote), and it uses a significantly different method from the game's.
I've also seen people mention locking teams, Shine's shuffle vote can be configured to do this. However, people will often leave and rejoin to try and unbalance the teams. The plugin does attempt to stop this by placing players onto the lower skill team if it can, but it can't always do this. This is more an attitude problem and is not something easily fixed automatically without seriously annoying players.
Another thing I've seen people mention is that the skills are not linear scaling. This may be so, but it doesn't affect team balancing at all. The balancing is done based on relative rankings comparing players against each other. No matter what (reasonable) transformation you apply, as long as it's the same to all players the relative rankings remain unchanged. You could remove data by rounding all skills off to a given precision (say, the nearest 500 points), which would perhaps change the sorting slightly, but it would still place high skill players as high skill, and low skill as low skill. The only possible difference would be the distribution of those near each other, which often matters little anyway as it's the extremes that cause the most issues.
At the end of the day, you can play with the numbers as much as you want, people will never be happy with it. There's only so much data you can contain in a single number, and a human is a whole lot more complex than that.
I was thinking of minimizing the sum of variances, but got stuck thinking that this is in most cases the exact opposite goal of balancing the teams, but minimizing the difference of variances seems like it could work much better.
Edit: I just had a thought. This is total overkill, but this could be a good theoretical framework. The goal of making the teams even is to make the teams as indistinguishable as possible. This concept, of making things indistinguishable, actually has a precise definition from information theory: Mutual information. So treat it as two random variables , the team selection in [0,1], and the skill values, which you can model as gaussians. The goal is to minimize the mutual information between the team selection variable, and the skill distribution variables. This metric should combine the difference of means and the difference of variances into one score to optimize for!
Can you talk a little more about what the algorithm does? The hive model is designed for skills to be added, which might be a bad model, but the scores are optimized assuming that's true, so adding them up so that the average is the same should be the right thing to do.
i don't know, oftentimes, you know the game is over. And it's really really really fucking over.
you're on 1 res, they've got every fucking lane covered by at least one guy.
you've been trying to sneak a skulk out for the better part of 5 minutes. They've been poking in, and you're pretty sure the other guys would notice an open tunnel/lack of defense.
and the guy in the other chair is almost certainly on point with his beacons because he's been on point with his scans so far.
and you can't get 2 skulks to kill a guy on an upgrade let alone into a fucking tunnel.
Yeah, when you're been outplayed, like really outplayed, there's a 5 percent chance of a comeback, and a 95 percent chance of dragging it out for another 10-15 minutes. just to lose while egg locked.
apologies, included another line somewhere for no reason
The best swap it finds that reduces the difference in team averages, and if possible, reduces the average in standard deviations (though it will allow the standard deviation to rise slightly if the difference in averages can be improved) is the one that is performed. It keeps at this until it can't find a swap that reduces the difference in average skill between the two teams anymore.
I can't say that it's the best thing ever, but from testing it produces consistent and (to the naked eye) correct enough teams. It's been a while since I did any formal statistics though so I'm sure that there's further improvements possible.
However, I will add that I have had games where we have the entire map as marines and for what ever reason we lose nanogrid on veil, and everybody wants to insta-quit. In spite of the fact that we have 3/3 jetpacks and 4+ rts. So the average pubber is not as good at judging when a game is over as they like to think in my experience.
Yeah, this is actually the detail I was curious about. How much difference in average skill is it willing to trade off for an improvement to the difference in standard deviation?
I got inspired by this and wrote up something that's a little more general. If I can get one of the programmers excited by it too then maybe we'll see it in the future.
https://docs.google.com/document/d/1EyXAzWDhE2YD37lDFmYcO2TvFztldsfseOH0-J8j2Tw/edit
The idea is based on what you are saying about making the mean and standard deviations as close as possible. It just defines that "as close as possible" a little more formally. The idea is that you model the skill distribution of the overall server as a normal distribution. Then you model the skill distribution of each team as a normal distribution. The goal in balancing the teams is to make the sum of KL Divergences between each team's skill distribution and the server's skill distribution as low as possible. That is, each team should have a skill distribution roughly the same as the overall distribution on the server.
I actually got my question answered quite well by the second person to post on this thread. So I don't understand what everyone else is babbling about. Something about how a better system is needed and that Hive Skill rating should be hidden. I think it works mighty fine right now.
Is it possible to have the system separate the rookies evenly before doing the standard shuffling that you described? I would like to see if it works better. Or let's say separate the people evenly with a hive score below a certain point and then do your average thing with the rest of the players.
A high round point is related to effective effort during gameplay. Players who can imprint effort and are successful in doing so (a metric already measured by the points) tend to be helpful to the team either by killing RTs/structures, by killing high life forms, by killing many low life forms which all eventually translates into a win.
This view captures rookies / avg players / good players on pub matches. I believe it can provide better rank discernment (improving all the benefits it brings) and also address the "categorization" issue.
Could the deviation from the avg team round points be used to weight (a non linear coefficient can avoid farming) the hive score points won/lost? Can it be tested?
Shine's hive score based shuffle actually already does this. Players with a hive score of 0 get shuffled using the fall-back method (by default last rounds k/a/d).
The fuck do you care for, you just leave the game and do whatever it is to make the hive score loss go away.
wooo, we got some salt there.
you're getting awfully close to a personal attack there.
Yes, the players with hive score of 0 get shuffled, but do they get spread evenly between both teams? I have noticed that all rookies end up on one team or most of the rookies are on one team after a shuffle (I think it was shuffle and not the other one). Are you saying they are spread evenly between both teams or they only factor into the average shuffle?
The skill is pretty well indicitive of who will win, upsets don't happen much, except in the last month or two where setting NEW rookies to 1000 while leaving old rookies at 0 completely FUBAR'ed the existing ratings. That rating change effectively meant that more skilled, but generally inexperienced players were significantly lower in hive rating than completely oblivious rookies, and your typical strong pub player has a rating so close to a fresh rookie (~1500 compared to 1000) that it doesn't make a significant impact on force even. The ratings might be able to even out eventually, but it'll be a long time going and every time a low rated "average" player comes back to play the game it will delay the equilibrium even more.
Anyway, outside of the RECENT changes, the problem with hive skill and force-even is that even if it's a close rating and "anyone's game", regardless of which side actually wins the end result is usually still a stomp. I don't think those games are actually "stacked", I think it's a product of rapid snowballing paired with pub players not having the know-how to turn the tables once it gets going. One team wins the first couple engagements by shear coincidence and secures enough of a lead that a pub team of of equal skill simply can't hope to get a leg up on them from that point on.
Didn't it start out that way? I know when I alt accounted it I needed to climb from 0 while some rookies were at 1000, that was around half a year ago?
Is it zero or close to zero? Rookies can luck into a win just by joining the correct team and won't always be at zero.
Thats actually a lie, pretty sure
New rookies don't get 1000 skill. The only "rookies" getting 1000 skill are people that already played before the hive reset that was like one year ago. So if an old player comes back, he will have 1000 skill. You can argue with that number, but the idea was that if somebody owns ns2 since several years, he is likely not a rookie with 0 skill and thus is started at 1000.
If you see any evidence that completely new rookies do indeed have 1000 skill, please link that users account. That would be a bug and is not intended.