value of rated games

[moboy78]

So, something I've heard on kgs and on forums like this before is that games from, say, 2 months ago don't mean as much to the rating system as a game from yesterday. Is this true? The reason I'm asking is because I recently played a game on kgs from which my opponent escaped, and it seemed as though he escapes often enough for me to get my win soon enough, but will that win still count as much it would've had my opponent resigned when the game was first being played? If so, why is that the case?

[DrStraw]

Why is it so important? Does what you learned from the game change over time? if not, then you have lost nothing. If it does, then rating will not retrieve it.

[moboy78]

Why is it so important? Does what you learned from the game change over time? if not, then you have lost nothing. If it does, then rating will not retrieve it.

It's not that it's important so much as it is that I'm curious about how the kgs rating system handle such things. I've heard different answers from various sources before, so I was hoping I could get a definitive answer to my question here on this forum, not another question.

[yoyoma]

Game weight has a half life of 15 days for weaker players, 45 days for stronger players. Details here:
http://senseis.xmp.net/?KGSRatingMath

[daal]

Game weight has a half life of 15 days for weaker players, 45 days for stronger players. Details here:
http://senseis.xmp.net/?KGSRatingMath

I think what the OP wants to know is when the game will count as being finished - as of the date the win was awarded, which in the case of a forfeit can be months after the game had been played, or as of the date of play, in which case it would have less weight with respect to his current rank.

[Mike Novack]

What might be argued here is that "real time" might not be the best way to "age discount" older data << assuming that game data should be age discounted >>

Number of games ago approximates time aging for players who are playing games at the average rate (per day, per week, etc.) but would also work for those playing at either a faster of slower rate.

[Mef]

What might be argued here is that "real time" might not be the best way to "age discount" older data << assuming that game data should be age discounted >>

Number of games ago approximates time aging for players who are playing games at the average rate (per day, per week, etc.) but would also work for those playing at either a faster of slower rate.

Weighting based on number of games played means that all weights will inherently be asymmetrical. This causes the rating system to become unstable.

[yoyoma]

Game weight has a half life of 15 days for weaker players, 45 days for stronger players. Details here:
http://senseis.xmp.net/?KGSRatingMath

I think what the OP wants to know is when the game will count as being finished - as of the date the win was awarded, which in the case of a forfeit can be months after the game had been played, or as of the date of play, in which case it would have less weight with respect to his current rank.

Ah right. I don't know about that.

[Mef]

Game weight has a half life of 15 days for weaker players, 45 days for stronger players. Details here:
http://senseis.xmp.net/?KGSRatingMath

I think what the OP wants to know is when the game will count as being finished - as of the date the win was awarded, which in the case of a forfeit can be months after the game had been played, or as of the date of play, in which case it would have less weight with respect to his current rank.

To my knowledge, games use the start date for ratings calculations and determination of when they are removed from the server.

[xed_over]

Why is it so important? Does what you learned from the game change over time? if not, then you have lost nothing. If it does, then rating will not retrieve it.

It's not that it's important so much as it is that I'm curious about how the kgs rating system handle such things. I've heard different answers from various sources before, so I was hoping I could get a definitive answer to my question here on this forum, not another question.

this interaction reminds me of when I'm at the bus stop, waiting for a bus...

often people will approach to ask me if i know what the schedule is, or when the next bus will arrive. They are usually shocked to discover that not only do I not know, but also that I don't seem to care to know.

Why should I care what the schedule is, or when the next bus will arrive? I still have to wait for it, either way. What possible difference could it make what time that will be?

[Bantari]

Why is it so important? Does what you learned from the game change over time? if not, then you have lost nothing. If it does, then rating will not retrieve it.

It's not that it's important so much as it is that I'm curious about how the kgs rating system handle such things. I've heard different answers from various sources before, so I was hoping I could get a definitive answer to my question here on this forum, not another question.

this interaction reminds me of when I'm at the bus stop, waiting for a bus...

often people will approach to ask me if i know what the schedule is, or when the next bus will arrive. They are usually shocked to discover that not only do I not know, but also that I don't seem to care to know.

Why should I care what the schedule is, or when the next bus will arrive? I still have to wait for it, either way. What possible difference could it make what time that will be?

Well, for example you might consider it more convenient to take a cab rather than wait for the bus, depending on how long the wait is.

In San Diego we have lines that go only in the morning and then again in the evening to accommodate people going to and from work but save the city money otherwise. So if you happen to need a bus at noon, it is nice to know there won't be one until 5pm. When I was living in Toronto, there were night lines which were going every 3 hours or so. It was nice to know if you just missed the bus and now have 3 hours wait in the cold or the next bus will come in 15 min.

I can probably come up with more examples why it is nice to know the schedule, but I think you can get my point.

[SmoothOper]

What if the difference were guaranteed to be 10-20 minutes(like on tygem where they lose if they don't reconnect in five minutes, or their clock runs down). Do you think you get any advantage from that 20 minutes of having a win?

[Mef]

will that win still count as much it would've had my opponent resigned when the game was first being played? If so, why is that the case?

Silly me, commented twice and ignored the original question...

Your kgs rank will be calculated the same. KGS ranks don't have a "memory" in the sense that Elo does (where you have your history summed as a number, that is shifted with each game), they are calculated by looking at all your have results in the last six months and finding a most likely rating to explain them.

[moboy78]

will that win still count as much it would've had my opponent resigned when the game was first being played? If so, why is that the case?

Silly me, commented twice and ignored the original question...

Your kgs rank will be calculated the same. KGS ranks don't have a "memory" in the sense that Elo does (where you have your history summed as a number, that is shifted with each game), they are calculated by looking at all your have results in the last six months and finding a most likely rating to explain them.

So, if I'm understanding you correctly, Mef, KGS only looks at the past 6 months of rated games that I've played and gives me a rating based on those results. The results of those rated games are given a value of either x, y, or z (x for if I've beaten a stronger opponent, y for a player of equal strength, and z for someone weaker) and uses the total value of x, y, or z to tabulate a rank that the system deems appropriate for me. The value of x, y, and z are constant throughout the 6 months that KGS uses them to estimate my rank, regardless of how far back in the 6 months that game was played.

[Mef]

will that win still count as much it would've had my opponent resigned when the game was first being played? If so, why is that the case?

Silly me, commented twice and ignored the original question...

Your kgs rank will be calculated the same. KGS ranks don't have a "memory" in the sense that Elo does (where you have your history summed as a number, that is shifted with each game), they are calculated by looking at all your have results in the last six months and finding a most likely rating to explain them.

So, if I'm understanding you correctly, Mef, KGS only looks at the past 6 months of rated games that I've played and gives me a rating based on those results. The results of those rated games are given a value of either x, y, or z (x for if I've beaten a stronger opponent, y for a player of equal strength, and z for someone weaker) and uses the total value of x, y, or z to tabulate a rank that the system deems appropriate for me. The value of x, y, and z are constant throughout the 6 months that KGS uses them to estimate my rank, regardless of how far back in the 6 months that game was played.

Kindof, but not quite. So what KGS does is it looks at the game result and the handicap for every rated game played by every player in the last six months. As you mentioned (and yoyoma expanded upon) it weights them based on the time the game was played (for the sake of this discussion when the game was started). Once it has this giant table of results (games, handicaps, and relative weights) it works out a simultaneous solution for every player on KGS for how likely it thinks any given player is to win in their next game against any other player (or to put this another way, "what is the most likely rating for every player given that we have this giant data table of results?" -- because of this KGS's algorithm is sometimes called a "maximum likelihood" algorithm). From that you get your rating differences. There are then some anchors who are used to set these relative strengths on an absolute scale.

So to make a long story longer -- KGS assigns you a rating based on all the rated games you have played in the last 6 months...and all the games your opponents have played in the last six months...and all the games their opponents have played in the last 6 months...and so on. The weight of the games decays exponentially with time (half-life of 15-45 days). It looks at both how you performed (win/loss) and how the game was handicapped (stones / komi / etc) and figures out the odds of winning between you and any other player on the server. The reason there is no "memory" is because this calculation is performed fresh each time using all the data, instead of using an incremental system (like Elo/Glicko/etc) where past data has been already lumped together.