A quick forward about the video: The presenter, Josh Menke, is not credited with having worked on or with HotS. Not though online seaching, or watching through all the credits.
So while this video may have impact on how HotS does matchmaking, J_osh Merke did not work on HotS matchmaking, so these exact design choices cannot be directly applied to HotS, not without some tinkering/thinking._
I will be attempting to keep my opinions out of this and merely focus on what is being discussed:
Timestamps:
2:00-2:20 about new and old players and possible matchmaking issues.
5:15-6:10 about how using simple stats can be problematic and an example
7:10-7:30, why finding skill quickly is important
8:30-8:40 elo versus unnnamed “state of the art” (likely a CoD system based on his game credits i could find).
9:20-9:50 how to judge how well calibrated a matchmaking system well, in terms of match prediction
11:45-12:10 about if a skill distribution is normal allows you to make fake players to test matchmaking/etc changes accurately.
12:30 graph of over 2 million players nearly fitting an expected normal distrubion. (Likely in a CoD game per Josh Merke’s credits)
12:40-12:50 most games he has worked on have had very near to a normal distribution in player skill rating.
12:30-12:45 About the Online Bayesian Ranking and being a good place to start for making a faster [than elo] ranking system.
13:50-14:15 about matchmaking and having issues with both large and small player bases.
14:55-15:30 about communicating to players what is happening with a search, and how lots of canceling and restarting searching is bad.
15:35-16:10 about how long until giving a worse match. brings up specifically shooter and moba, with vague guesstimates of time.
16:30-16:40 about putting monetary value on each player for each match, Josh at the time did not think that anyone knows how to do that (possibly still does think no one knows?)
18:50-19:30 about matchmaking over time and current “Global Optimizer”.
20:05-20:45 Matchmakers and complicated, using fake player matches to test them (debugging) and testing ranking.
21:05-22:15- tight matchmaking versus random matchmaking
22:45-23:10 how matchmaking correctly lets you create skill depth in game.
23:35-24:35 about how to make groups and playing with friends work. (please note, from what HotS team has said about matchmaking in groups, it sounds like HotS does not use the exact “solution” that Josh Merke presents here)
25:10-25:40 issues of people playing with friends followed by solo
26:00-37:00 about three different ways to reveal a players skill to the player.
timestamps in this:
—26:10-28:05 about loading screens and ranking systems
—30:05-30:50 more about loading screens and matchmaking
—32:00-32:15 probably the kind of system that HotS uses
—32:20-32:45 about hiding versus showing true skill
—33:10-33:20 we can simulate these ranks
—33:50-34:10 pure skill system for matchmaking example
—34:50-35:10 why matchmaking on rank can have issues
—35:25-35:40 how you can get good matchmaking with rank and skill
-37:10-42:20 about how Josh Merke likes to do things (I split out some sub points that relate to HotS and how his work may have influeneced HotS design team)
—>>>38:10-38:45 about placement matches
—>>>39:15-39:45 about broad ranks and winrates between players of different ranks against each other
—>>> 41:00-41:35 how Josh Merke likes to distribute rankings of players
Or i’ll do it now while i wait for some large files to download… please note i am only listening the questions, although the times given include the answer
Q1: 42:30-44:20 about server sending lots of games versus sending one game to the client. also about player churn between these two.
Q2: 43:25-46:00 have you ever been privy to any goals or motivation for not having a 50/50 win-loss goal?
Q3: 46:05-47:20 The skil system you talked about based on win/losses, similar to elo, have you though about using more in game metrics to rank players, like about different roles and position in moba’s?
Q4: 47:25-48:30 experience on small player games–how large does the playerbase need to be to make a skill based matchmaking system make sense?
Q5: 48:30-50:30 What are your thoughts on decay, about skill being lowered the longer someone does not play. such as if someone doesn’t play a moba for a month the meta changes, even if the individual player’s skill does not.
Q6: 50:35-51:10 It seems like a player’s skill variance would account for someone not playing for a while, so why decay the rating? Isn’t the math supposed to handle that?
Q7: 51:15-50:40 When you matchmake, you do it on the chunks right? Not granularly? Like if you’re going to match anyone in silver, you just match them against anyone in silver–or do you do 2200 versus 2300?
Q8: 51:40-52:25 The popularity of the hybrid system, I feel like LoL and these games are like going to go through this progression based system, and then transfer you over when we’re sure you’re ready, because we don’t want you to drop or bounce because we’re seeing this number go down. Can you talk a little bit more about that, because you were saying “oh, just go skill” which is obviously a seductive thing if you’re making an esport kind of game, but you’re just worried about people just bouncing from the game.
Q9: 52:30-53:40 Do you have critical mass issues in games? how you weigh the tradeoffs between giving people matches that are not great for them to wait long, versus making them wait to get a better match, and how you balance those factors?
Q10: 53:45-55:15 So for some games that have different roles, players might have different skill levels, depending on what they end up doing. Any thoughts on hwo you can matchmake and determine skill based on the roll they end up playing? both about situations where the role is pre-chosen and is post-chosen of mathmaking
Q11: 55:20-56:50 What are the dynamics of skill over time? So if you start off with a given skill distribution, you showed high skill players at a constant skill, but bad players skill creeping up over time earlier. Do you see the opposite effect where everyone is normally distributed that the variance of skill narrows, or feels constant over a game?
Q12: 56:55-58:25 How do you think ranked reset effect hybrid models? Once a year you reset, and good players get a higher rank to start, and lower players a lower rank. For example in HS, players reset every month. Players get caught up in the middle after reset and causing discomfort in bad players.
Q13: 58:30-59:25 But do you think maybe you would have, just a larger difference in reset? Say if i was a 11 player in HS who resets to rank 15, if i set them at rank 12 instead, where they’re unlikely to meet up with newer players, or bad players, do you think there are different ways like that?
Q14: 59:30-1:00:40 How would you deal with players who might intentionally lose to “smurf” to try to play against worse players?