Navigation Content Footer

Dropper Detection Explained

06/19/2011 — Next to the players' names, colored bars show up that signify whether it is worth the while playing these people. But how are these values computed?

So, what are droppers? When playing on FIBS, sooner or later you'll learn, that your matches are interrupted. There are several reasons for this.

One of them are simply network problems. This happens quite often with players using mobile devices. Consequently, some people on FIBS deny playing anybody using a mobile phone.

The “classical” droppers, however, are people that cut their connection to FIBS without warning, as soon as it becomes clear that they will lose their match. The motivation for that neurotic behavior is obviously to manipulate their rating. When you never lose, your rating will never decrease.

Either way, it is a nuisance and several strategies have been developed avoiding this. One of them is to refuse playing against mobile users. Another option is to use RepBotNG. RepBotNG is a non-playing computer user that collects complaints or vouches, and computes the “repuatation” of other players based on this data. This is technically awkward and I personally find that idea a little problematic from an ethical point of view.

The so-called “saved count” also gives information about the reliability of other players. Whenever a match is interrupted, FIBS stores the current state for a maximum of 14 days. Gibbon retrieves this information automatically for you, whenever you invite somebody or get invited. Alternatively, you can query that number manually (by typing „show savedcount OTHERPLAYER“ into the server console). The problem with that number is that the saved count increases for both opponents, when a match is interrupted. In other words, the “victim” and the “culprit” suffer from the same penalty.

Gibbon's strategy is extremely simple yet efficient. Actually, the data needed for estimating other players' reliability is available. It just has to be collected: FIBS sends a notification to all users, whenever a match starts, stops prematurely, or is resumed. This data just has to be processed and interpreted. And the big advantage is that it is always clear who was responsible for the premature termination of a match.

Gibbon simply monitors what happens on the server, and notices (and remembers) which players drop their connection during a match, and which players finish their matches in aregular manner. From this data, the software computes indicators that get visualized by red, orange, yellow, or green bards with varying length. Red signifies “keep away“, green is an indicator for very reliable opponents, orange and yellow are used for the neither fish nor fowl area.

For each user on FIBS two values get computed, the “reliability“ and the “confidence„ for the reliability. Players with a strong affinity to dropping their matches have a low reliability value, those that finish them have a high reliability. For very active people, the confidence value is high (because Gibbon has a lot of data available for them) and for less active players, or players that are still unknown, the confidence will be low. The reliability is visualized by the color (red, orange, yellow, green), the confidence by the length of that colored bar.

In detail, it goes like this: When FIBS sends a notification that a user has dropped the connection, Gibbon checks whether that player was participating in a match at that time. If this is the case, Gibbon penalizes the drop with -1.0 points, otherwise nothing happens..

When FIBS sends a notification about a regular end of a match, both players get a bonus of 1.0 points. By the way, this happens a lot more often than the former case.

And what happens, when Gibbon receives a notification about a resumed match? It depends on time time that the match was interrupted. If that happenedrted before the current session, there is no way to find out who of the two opponents was the dropper and who was the “droppee”. Gibbon will then ignore this resume. But when the culprit is known because the match was interrupted during the current session, the dropper gets a bonus of 1.5 points. This bonus is not too high! We will see that right away.

The reliability of a player is simply the arithmetic mean (the average value ...) of the bonuses and maluses. The confidence is the number of registered events for that player. A long bar indicates an acdtive player, the color of the bar indicates the registered behavior.

Let's look at some typical scenarios. The normal case are players that start their matches, and play them to the end. Each of these matches will result in a bonus of 1.0 points for both players. The average (and the maximum value!) is therefore exactly 1.0. The reliability bar for such FIBSters has a lush green and quickly grows to the maximum length.

A typical dropper cuts the connection as soon as the match is clear to be lost. In the long run, Gibbon will measure a value hover around 0 for their reliability. For them, each victory will “earn” them 1.0 reliability points, each (potential) loss is recorded as -1.0 points. Even extremely good backgammon players still lose their matches from time to time, and therefore the reliability value for notorious droppers will always be low enough to distinguish them from the nice guys.

A third group of people has an instable connection to FIBS. Maybe they are traveling through areas with varying network coverage, or they have private reasons to interrupt their matches frequently. Let's take a player that is willing to player but has network problems, and therefore gets kicked out of FIBS very often.

A match that is played in one single session results in 1.0 reliability points. If the match was interrupted once, then the dropper will get -1.0 points as a penalty for the drop. If he resumes, he gets +1.5 (!) points, when he finishes the match regularly, a third event with 1.0 points is stored. In sum this is 1.5 points (-1.0 + 1.5 + 1.0), the average is 0.5 points (1.5 / 3 = 0.5). That means that the average value can never exceed 0.5 points for an interrupted match, and that value is already displayed as dark red in Gibbon.

A match that was interrupted twice results in 0.4 reliability points, 2 x -1.0 for dropping twice, 2 x 1.5 for resuming twice, +1.0 for finishing, everything divided by 5 events gives 0.4 points. If a match was interrupted an infinite number of times, the average value converges to 0.25 points.

Hence, the relevant interval for estimating other players' reliability is from 0.25 to 1.0. All values below 0.25 are completely out of question.

The algorithm used by Gibbon knows one exception, and that is for bots that play matches. Drops against computer players are treated exactly the same way as between humans. However, resumes are treated differently. In that case, the penalty (-1.0 points) given for the interruption is completely deleted from the database. The human player is treated as if the interruption has never happened. The rationale for this is that bots are not bugged by interrupted matches. And therefore the behavior is acceptable, as long as the match gets resumed later.

Currently players with a reliability value greater or equal to 0.95 are marked with a green bar. Players with at least 0.85 points still get an orange “acceptable“ bar. Those with at least 0.65 already have an orange bar, and players with reliability values below 0.65 have a red bar. The maxium value for the confidence is 10. For higher confidence values, the length of the bar no longer grows. Events older than 100 days are deleted automatically. That means that Gibbon forgets bad (and good!) behavior after some time.

The strategy used by Gibbon is simple but very efficient. When you are looking for a potential opponent, simply look out for those with a wide green bar. In the very beginning, of course, all bars are white (empty) but that will change rapidly, when you are logged in a couple of minutes.

People have complained that Gibbon completely ignores the motivation for interrupting a match. Classical droppers are treated exactly the same as people that suffer from an instable network connection, or players that frequently get confronted with sudden reguests by their bosses, children, parents, or ... you name it. However, Gibbon does not evaluate the character of other people. How could it? It is a stupid computer program. It simply tries to give the user an idea of how probable it is that a user with another player on the server can be played without interruptions.

If you know the reason why some particular player has a red bar, and you still think this is a good guy, go ahead! Nobody hinders you. But otherwise, Gibbon equally protects you to a certain degree from neuroticis with a FIBS rating fetish and people that are “soooooo“ busy that they can often not play their matches to the end.

blog comments powered by Disqus