
- This picture shows the relative strength of players. For example: if you are at the height 0.2 (indicated at the left) then your killratio versus players at height 0.1 is estimated to be 2.
- When you click on a player you see additional data, among which the basic one-on-one all the other data is derived from. This usually doesnt work with older browsers or internet explorer.
- It is updated every few hours.
- It is currently based mostly on data from mofo, about a dozen other servers (still all on the same machine) participate too.
- Only kills between registered players are counted, TK's are ignored.
- The calculation is done using a genetic algorithm, the ranking wont be perfect but it is probably a good estimation. I dont even know how the space of solutions look like: Are there local optima? Is the solution stuck in one of them?
- The basic assumption is made that the strength relation is transitive: if A shoots B 5 times and gets shot only once, and B shoots C 3 times and gets shot only once, then A supposedly would shoot C 15 times to be shot only once.
(This isnt totally right in the usual multiplayer situation.) This is the main selection criterium for the genetic algorithm.
- There is a second criterium with very low importance: it pushes players down on the ranking. It's value is too small too interfere with the main criterium, it becomes active when the main criterium is indecisive. A sizeable group of players with rank 0 are not shown in the image. (they never shot a registered player.)
- It also assumes players do their best.
- Because the data comes mostly from mofo, the ranking is limited to the playstyle and player population there. (More servers should record data, when that becomes possible.)
- It uses the notion of "connectedness". The more data is available about a player the more connected he gets. The connectedness of the opponents (and their opponents and so on) also matters.
- Players with a low connectedness will have a very opaque label on the image, the players who are well connected have a well-visible label.
- Your score against a well connected player means more compared to your score against a less connected player.
- The algorithm used contains no (or barely any) arbitrary values, it should be mathematically honest.
- If you don't like your ranking, this should be enough data to figure out why you got it and against who you have to play differently to change it.
- If you want to game the system you will need to sacrifice at least one dummy account (that can be noticed on the site by anyone). And you wont be able to play regularly otherwise the system will correct your rank to your play. To avoid creating noticable records in the database you could use lots of dummy accounts, but their relative lack of connectedness to the rest of the graph will be found. (Actually a part of me wants someone to game the system so that i could toy more with graph-algorithms.)
- Know that the image uses a logarithmic scale. Whatever rank the median player has, the logarithmic scale is chosen in such a way that he goes to the center of the image.
On the image, a height difference on top corresponds to a larger difference in rank compared to the same height difference at the bottom.
- The rank itself means nothing. It only means something relative to the rank of other players. Comparing to the rank of the median says the most.
-
This thing is not very dynamic like a dayly score. If you have a long history of playing at level x and then your play changes ... it might take you a while to move in the image.
- A xhtml+svg enabled browser can show the error curve for the selected player. For the domain of ranking values [0,1] it plots the error value, scaled in such as way as to give an idea about the curvature. It is not clear if the rank the player has, really needs to have the lowest error because the ranks of others have to be considered too.