Learn how matchmaking works on Roblox — Transcript

Full transcript

1. 0:00In this video, SuperPAT77, a Principal Software  Engineer at Roblox, teaches you how matchmaking
2. 0:06works in experiences, including a breakdown of  the different signals used to match players into
3. 0:10eligible servers. With the launch of custom  matchmaking, it's important to understand how
4. 0:15our matchmaking algorithm works so that you  can modify it to match players according to
5. 0:19the unique needs of your experiences. For more  information on the matchmaking process, as well
6. 0:24as how to customize it for your experiences,  check out the links in the description!
7. 0:33Let's dive into the details of matchmaking.  Let's say a player finds your game, they go
8. 0:37click the big blue play button, and matchmaking  is going to go put them into a server. But how did
9. 0:43we actually decide that this was the server to go  put them in instead of one of these other servers
10. 0:48over here? We have two steps for our process.  The first step is what we call eligibility.
11. 0:54We need to find servers that are eligible  for the user to join. That means filtering
12. 0:58out servers that aren't running the right place  or filtering out servers that are already full
13. 1:02like this middle server over here. Or maybe  you've just launched an update for your game,
14. 1:06and we need to make sure players don't join  servers that we're about to go shut down. We
15. 1:10have a series of eligibility checks for all of our  instances, and once we run through all of that,
16. 1:14we have a list of instances that are eligible for  the user to join. The next step is what we call
17. 1:20scoring. We have a function that scores every  single one of our servers, and whichever one is
18. 1:25the highest scoring server, that's the one that  the user is going to join. With today's launch of
19. 1:31custom matchmaking, you will be able to change  how scoring works for your experience so that
20. 1:35you can decide whether this player should go  to the top server or the bottom server. Now,
21. 1:41let's talk through how scoring works. We  use an algorithm that we call weighted sum.
22. 1:48Weighted sum means that for every server we have  a list of signals like how full the server is, how
23. 1:54good your latency is. We're going to go convert  this raw datas into a value between 0 and 1, we're
24. 2:00going to multiply it by a weight, we're going  to sum all of those weighted signals together,
25. 2:04and that becomes our weighted sum. Let's walk  through an example now. We have several signals
26. 2:11during matchmaking, but to keep things simple,  we're just going to use two signals: we're going
27. 2:17to use one signal for occupancy and we're going  to use one signal for latency. Now, I mentioned
28. 2:25every signal has two parts: a value and a weight.  Let's talk about the value first. For the value,
29. 2:32we want to be able to convert the raw signal data  into some number between 0 and 1. We want to do
30. 2:37this so that no matter how different the data is  across these signals, we have numbers that we can
31. 2:41compare across those different signals. So how do  we do this for occupancy? It's pretty simple. We
32. 2:47take the number of players, and we divide it by  the capacity, the total number of players that
33. 2:54could be inside of that server, and that gets us  some number between 0 and 1. For this first server
34. 2:59up here, we would just take 4 / 5 and we would  have a signal value of 0.8. For this second server
35. 3:06down here, we would take 2 / 5 and we would have a  signal value of 0.4. But what about latency? What
36. 3:13do we divide 100 by? What do we divide 50 by? How  do we know how good 50 is versus 100? We have to
37. 3:21pick a number, and it has to be consistent across  all of our servers so that we can have the same
38. 3:25comparison across the board, and we picked the  number 250. Now, this isn't to say that you can't
39. 3:34join a server that has a ping higher than 250, but  it does mean that beyond that point, we might as
40. 3:38well use a value of zero. It's no longer relevant  for scoring. We want to make sure that we keep
41. 3:43this between 0 and 1 so if it is greater than 250,  we'll need to make sure that we constrain that.
42. 3:51And then, well, this isn't quite right. If you  have really good ping where you have, let's say
43. 3:55only 10 milliseconds of ping, the value is going  to be close to 0. So what we actually want to do
44. 3:59is take the inverse of this number, and now we're  ready to start converting our latency data into
45. 4:08signal values. So let's do that! For this first  server over here, we would have 1 - 100 over 250,
46. 4:15which is equal to 0.6. And for our second server,  we'd have 1 - 50 over 250, which is 0.8. Now, we
47. 4:27could stop here. This is enough for us to compute  a score for all of our instances, and let's see
48. 4:31what that would look like. If we were to just add  these numbers together, these values together,
49. 4:38we'd get a final score of 1.4 for our first  instance and 1.2 for our second instance. And
50. 4:52that means just by doing this, we would send the  player over to this top server over here. But what
51. 4:57if that's not the right trade-off? What if that's  not what we wanted to do? What if we say, "Hey,
52. 5:02you know it's great that this server has a lot of  players, but if the player joins the server their
53. 5:07ping is never going to get better. Whereas if they  join this server down here, maybe it's okay that
54. 5:12there's less players, more could join later,  I care more to make sure that they have good
55. 5:15ping." So this is where we introduce the concept  of weights. Now, weight can be any number between
56. 5:220 and infinity. It doesn't matter, and we can pick  some values. So let's say we picked a value of 1
57. 5:28for the occupancy and 3 for the latency, and  let's see how that changes the outcome. Now,
58. 5:33we're going to multiply our occupancy signal  value by 1, and our latency signal value by 3,
59. 5:41and what we find after doing that is that now  we have a score for the first instance of 2.6,
60. 5:48and a score for the second instance of 2.8 . And  so now by changing the weights, without changing
61. 5:53any of the data, the player would be joined  to this bottom server over here. This is how
62. 5:59the weighted sum algorithm works, and this is how  matchmaking works every time user clicks the play
63. 6:03button. We have more signals, those signals have  their own transformation functions to keep those
64. 6:09within those values, and they all have different  weights. And now you, as a developer, have the
65. 6:13ability to define your own signals using your data  and your own weights for our signals and for your
66. 6:19own signals. We hope you get a lot of value out  of this new feature, and we hope you're able to
67. 6:23customize matchmaking to be the best for your  experience's player base. Thanks for listening!