21 replies to this topic

[Sentrosii]

Hi,

 

Forgive me if this has already been asked but, how does silent calculate the k/d number?

Kill ratio is addicted to weapon which we are using?

For example, kill ratio is different to each weapon, like, i dont know, knife x2, mp40 x 1 and sten is x0.2?

How does it calculate? Someone could help me out with this?

How does it calculate for landmines, arts, granate, mp40/thomson, sten, fg42 etc?

Please its really important for me.

 

And how does it work with PRW?

 

Thanks for answer.

[gaoesa]

KR is calculation uses distance between the players,the average distance between players on the server and the old KR values of the players.

 

The PRW is more complicated and that is better explained by the pdf. But I don't find it right now.

 

Do note that both the implementation and the documenation is from the ETPub which was the base we started. But the documentation of the KR is wrong about taking weapons into account.

[BECK]

Isn't there something also about the skill level (probably PRW) of the player who is killing verses the player who is killed?  For example, if a really good player kills a noob, the change in the rating is minimal, whereas the noob killing the good player, the calculation favors this more?

 

I've noticed that in the spree stats for each map, there is rating variance and things like that that are stored.  I'm assuming this data is used in the calculation somewhere, which may explain why some people complain that KR is so weird on some servers and doesn't make sense?

[Spyhawk]

Isn't there something also about the skill level (probably PRW) of the player who is killing verses the player who is killed?  For example, if a really good player kills a noob, the change in the rating is minimal, whereas the noob killing the good player, the calculation favors this more?

 

I've noticed that in the spree stats for each map, there is rating variance and things like that that are stored.  I'm assuming this data is used in the calculation somewhere, which may explain why some people complain that KR is so weird on some servers and doesn't make sense?

 

Yes, ETPub's KR takes skill of killed player into account. Making "easy kills" on noobs won't improve your KR much, but killing a much better player than you will improve your KR much more.

 

KR is independent of PRW, and doesn't really make sense in a multiplayer game since it takes into account only the attacker and the victim, much like an independent 1vs1 game. But PRW is quite interesting imho. PRW also (optionally) takes the difficulty of the map into account.

 

If you're interested in knowing more, look for J.Menke's documentation: "A Bradley-Terry Artificial Neural Network Model for Individual

Ratings in Group Competitions", and "Hierarchical Models for Estimating Individual Ratings from Group Competitions" (the later being an updated version). Strong background in mathematics/statistics recommended.

Edited by Spyhawk, 07 January 2014 - 04:36 PM.

[gaoesa]

I think the terms are little mixed here. The PRW doesn't count kills at all. It is counted from winning the matches and the player playtimes in the teams. KR is the only method that rates the kills. From these two, I have found that KR is much more accurate. The PRW seems to give a lot of weird results. Though the theory sounds good. But there seem to be too many variables that affect the outcome of the maps for this to become accurate in this specific implementation. It's not surprising that the KR is giving reasonable results. All of the game revolves around kills afterall. Of course, both of these methods require that the teams are playing for the objective and try to win the map.

[Spyhawk]

Thx, you are right gaoesa. I changed my last post (see the bold) - I was thinking about KR while I wrote PRW, my bad.

 

I'm personally not convinced that KR is accurate. It is still better than nothing, but it won't reflect action where many players are involved (ie, you're fighting an enemy, almost killed him, but a teammate will "finish" the kill). Only the victim and the player that makes the kill are taken into account.

 

 

 

All of the game revolves around kills afterall. Of course, both of these methods require that the teams are playing for the objective and try to win the map.

 

Not really - as you said, PRW is looking only at the outcome of a game while KR looks only to individual 1vs1 combat, so if you are the best killer on a server but a poor objective player, you might end up with a high KR, but a low PRW.

[gaoesa]

You're right. KR does not provide accurate information. There are lot of ways how to play to get cheap kills and high KR. Also, players are inconsistent on how they perform in the game even if they are just playing. What I meant in that quote though, was that in my opinion, making kills is the most important thing in an FPS game, if the players in both teams are doing the killing to attack and defend the objectives i.e. to win the map and not for personal stats.

[Spyhawk]

Btw, I recently found a bug in the time played in each team values  in etpub (mapAxisTime / mapAlliesTime). It is not computed when players are in limbo, so the more a player spend in limbo, the less accurate these values are. Consequently, the PRW can't be accurate either. Chances are that this bug is still present in latest silent.

Edited by Spyhawk, 15 January 2014 - 11:07 AM.

[TheSilencerPL]

Chances are that this bug is still present in latest silent.

 

It's still present. Thanks for the info, we will think this over and fix it if needed.

[TheSilencerPL]

Btw, I recently found a bug in the time played in each team values  in etpub (mapAxisTime / mapAlliesTime). It is not computed when players are in limbo, so the more a player spend in limbo, the less accurate these values are. Consequently, the PRW can't be accurate either. Chances are that this bug is still present in latest silent.

 

On second thought, I think this is the way it should be as it takes under consideration only the real play time. Otherwise, if the player was not helping his team (by staying in limbo or spec) the PRW would not be reflecting the real situation.

[Spyhawk]

On second thought, I think this is the way it should be as it takes under consideration only the real play time. Otherwise, if the player was not helping his team (by staying in limbo or spec) the PRW would not be reflecting the real situation.

 

At least, this behaviour should be consistent with dead players and players in limbo. Both a player that is waiting for a medic or a player that is forcefully sent to limbo contribute nothing to the gameplay, yet only the time of players in limbo isn't taken into account. This would also implies that players in limbo aren't part of a team, which doesn't really make sense.

 

This clearly seems to be an unwanted bug (see the return right above the mapAxisTime/mapAlliesTime computation). Beside this, the behavior isn't documented anywhere - nor in a comment in the code, nor anywhere in the documentation.

 

It's hard to evaluate how much this bug influence the PWR rating in average. Maybe by a few percent only, but this is still a systematic error.

Edited by Spyhawk, 17 January 2014 - 11:24 AM.

[Sentrosii]

Thanks a lot for help. Now I trying to understand how does it work.

 

I know kr is calculating by distance between players and theme kill ratio.

 

Anyway... does weapons is important to change kr?

Is it difference beetwen kills by arts, landamines, granats or for example beetwen sten and mp40?

[gaoesa]

Weapon types have no direct influence to the calculation. With different weapons the distances are different so they can have indirect effect.

[Sentrosii]

OK but what about landmines, arts or granates?

How does it includes in KR or PRW?

[belstgut]

Maybe take a look at the TrueSkill Algorithm by microsoft for upcoming versions. TrueSkill also provides logic for matchmaking and balancing teams.

 

http://www.moserware...your-skill.html

http://dl.dropbox.co...d TrueSkill.pdf

C# Implementation: https://github.com/moserware/Skills

Online Calculator: http://research.micr...alculators.aspx

[Sentrosii]

Wow, this looks really nice.

Thanks for that links.

 

OK so now i have to read this all.

[belstgut]

there is no point in reading this unless you really know your statistics and propability mathematics

[Spyhawk]

there is no point in reading this unless you really know your statistics and propability mathematics

 

No necessarily. The MoserWare article you linked to introduces the needed basics in quite a simple way, unlike the original TrueSkill paper.

I don't think TrueSkill brings much to the table as the mostly similar ETPub PRW is already included, but it could be a nice addition for others mods. I've been working on implementing it for Legacy for some time now. The code is not finished yet but you can expect to see this in ET soon (most of the code is already written, but others stuff have higher priority right now).

[belstgut]

yeh i mainly meant the pdf

 

nice to hear that you are implementing this

 

can we also expect some kind of matchmaking system? that would be awesome ^^

[Spyhawk]

I was specifically referring to the original TrueSkill PDF. The Moserware PDF is much easier to read.

 

As a first step, I'd like to implement only the TS algorithm as it is currently done with ETPub Player Rating, with features like winning probability and optional match balancing.

 

Later, match making might be implemented too, depending on the difficulty in doing so. Afaik, this would require some features in a reworked masterlist code, so this shouldn't be done in mods, but in an upgraded server infrastructure (such as what is planned with the ET:Legacy project).