标题: DOTA2天梯排位系统即将到来 [打印本页] 作者: 第八铜人 时间: 2013-12-8 10:13 AM 标题: DOTA2天梯排位系统即将到来 Over the past several months we’ve been working on improving matchmaking. In this post we’d like to share with you where matchmaking currently stands and give you a sneak peek on an upcoming matchmaking feature.
Ranked Matchmaking is Coming
The next major update will add a ranked matchmaking feature to the game. This mode is aimed at experienced players who want to play in a more competitive environment and know their matchmaking rating (MMR). Dota 2 matchmaking has always calculated MMR and used it to form matches; in ranked matchmaking we make that MMR visible.
作者: 第八铜人 时间: 2013-12-8 10:13 AM
Your Matchmaking Rating (MMR)
Dota 2 uses standard techniques to quantify and track player skill. We assign each player an MMR, which is a summary metric that quantifies your skill at Dota 2. After each match, we update your MMR based on what happened in that match. In general, when you win, your MMR will go up, and when you lose, your MMR will go down. Win/loss is the primary criteria used to update MMR, but individual performance also plays a role, especially when our uncertainty about your MMR is high. It is possible for an individual MMR to increase after a loss or decrease after a win, but in general the winning team’s average MMR will increase and the losing team’s MMR will decrease.
We also track our uncertainty about your MMR. New accounts and those playing in Ranked Matchmaking for the first time have high uncertainty. Higher uncertainty allows larger adjustments after each match, and lower uncertainty leads to smaller adjustments. Together, the MMR and uncertainty can be interpreted as a probability distribution of performance in your next game; the MMR itself serves as the mean of this distribution and the uncertainty is its standard deviation. If the match outcomes (both the win/loss and individual performance) repeatedly match our expectations, the uncertainty tends to decrease until it reaches a floor. A surprising match outcome will tend to cause an increase in uncertainty.
We actually track a total of four MMRs for each player:
Normal matchmaking, queuing solo
Normal matchmaking, queuing with a party
Ranked matchmaking, queuing solo
Ranked matchmaking, queuing with a party
同时，我们也会追踪你的MMR，新账号和第一次玩RM的账号具有高的不稳定度，高不稳定度的账号在每场比赛之后，对于分数的调整会更大，同时，低不稳定度的账号调整会小一些。
我们将会纪录以下四几种MMR数值：
1、 普通匹配(NORMALMATCHMAKING以后称为NM)，单人匹配
2、 NM，多人匹配
3、 RM，单人匹配
4、 RM，多人匹配作者: 第八铜人 时间: 2013-12-8 10:15 AM
Each of the two ranked MMRs has its own calibration period. Under certain circumstances, we may need to reactivate calibration, if we think the MMR is inaccurate.
To give you a feel for the range of MMR, below are some MMRs corresponding to various percentiles.
这个比例并是从以前的NM中得来的，我们还不知道在RM中，这些分数的分布会是怎样呢，但是我们也认为他会有所不同，因为进入RM游戏中的玩家将会是更为高手，更富经验的玩家群体，我们也预料到了，玩家在RM游戏中，将会和他们在NM中的表现不太一样。 作者: 第八铜人 时间: 2013-12-8 10:17 AM
Note that this distribution is from normal matchmaking. We don’t know yet what the distribution will be in ranked matchmaking, but we expect it to be different. The players who participate in ranked matchmaking will be more skilled, more experienced players. We anticipate that any given player will have different expectations and play the game differently in ranked matchmaking compared to normal matchmaking.
What Makes a Good Match?
The ultimate goal of automated matchmaking in Dota 2 is for players to enjoy the game. The matchmaker seeks matches with the following properties (listed in no particular order):
1)The teams are balanced. (Each team has a 50% chance to win.)
2)The discrepancy in skill between the most and least skilled player in the match is minimized.
3)This is related to team balance, but not the same thing.
4)The discrepancy between experience (measured by the number of games played) between the least experienced player and the most experienced player is minimized. More on this below.
5)The highest skill Radiant player should be close to the same skill as the highest skill Dire player.
6)Each team contains about the same number of parties. For example, the matchmaker tries to avoid matching a party of 5 against against 5 individual players.
7)Players’ language preferences contains a common language. Lack of a common language among teammates’ language preferences is strongly avoided. Lack of a common language across the whole match is also avoided, but less strongly.
8)Wait times shouldn’t be too long.
什么会早就一场好游戏?
匹配系统的终极目标是让玩家享受游戏的过程，匹配系统是通过以下的方法来搜寻一场游戏的：
1、队伍之间是平衡的，每个队伍都有50%的概率能赢。
2、在一个队伍之中，最高手和最菜鸟的玩家之间的差距是最小化的，这和队伍之间的平衡
3、是相似的，但是不是同一种东西。
4、一个队伍中，最高场次的玩家和最低场次的玩家的差距是最小化的
5、双方的最高水平玩家的水平是一致的，(大酒神没法从零单排了)
6、双方的黑店玩家的数目是尽可能一致的，匹配系统已经尽量在避免五个黑店匹配五个单人玩家了。
7、匹配系统将会尽可能希望将同种语言的玩家匹配到一起，队友之间所使用的语言不一样是尽可能不免的
8、等待时间将会尽可能不要太长(也就是说，匹配时间太长的话，匹配系统就会缩小这些限制) 作者: 第八铜人 时间: 2013-12-8 10:18 AM
The matchmaker seldom achieves all of those goals perfectly. For any potential match, the matchmaker assigns a quality score for each of the criteria above and then takes a weighted average. When the overall quality score exceeds a threshold, the match is considered “good enough” and the match is formed. We’re constantly experimenting with different match criteria and how to prioritize them.
The matchmaker does not directly try to achieve any particular win rate for players. However, we do try to ensure that each team has a 50% chance of winning in any given match. (This is criteria #1 in the listed above.) We do not examine individual win / loss streaks or try to end them. However, if you are on a winning streak, in general your MMR is probably rising, which will tend to cause you to be matched with higher skilled opponents and teammates. Win rate is not a meaningful measure of player skill.
Win count is also not useful as indicator of skill, and the matchmaker does not use it for that purpose. We do try to group players by their level of experience (criteria #3 in the list above), primarily because we have found that players at the same skill level but different experience level differ in their expectations of how the game is to be played. Our measurement of “experience” for matchmaking purposes is an approximately logarithmic function of the number of games played. The difference in experience between 40 games and 120 games is considered to be about the same as the difference between 120 games and 280.
You can visualize the impact of goals #2 and #3 with a chart where number of games played is the horizontal axis and MMR is the vertical axis. If two players are close together in the diagram, they are considered good candidates to put into a match together. Players who are far apart are considered a poor match. The typical career trajectory of a player new to Dota 2 as he gains experience and moves towards the right is to gradually move upwards as their skill increases. When skilled players create new accounts, they follow a bit different trajectory. Their MMR rises relatively quickly, placing them into the top lefthand corner of the diagram, where they will be matched with other players whose skill is high relative to their experience level.
对于一个新手玩家，我们将会让他的分数随着他的游戏过程慢慢上升，而对于一个高手新建了一个账号，那么我们将会使他们的分数快速上升，以便使他们匹配到恰当水平的游戏作者: 第八铜人 时间: 2013-12-8 10:19 AM
What About Parties?
When parties are involved, things get a bit more complicated. Parties often contain players with a wide discrepancy in skill and experience. For the purposes of measuring the goodness-of-fit criteria listed as #2 and #3 above, the matchmaker assigns each party aggregate skill and experience numbers. It is these party numbers that are used rather than the individual. In general, when a party with a wide skill range is matched with a solo player, the solo player will have skill and experience near the average of the party. If you notice that one player seems to be significantly less skilled than the other players in the match, it is very likely that they are partied with a high skilled player.
Also, when players are in a party, they typically perform better than players of equivalent skill who don’t know each other. We account for this in two ways. First, we track your skill when queuing alone separately from when queuing in a party. Second, we adjust the effective MMRs based on the number of players in the party and the distribution of skill within the party.
Here’s an example match that was formed today that demonstrates both of these principles in action.
