ratings are currently displayed as x(y) and sorted by x-y. I'd suggest displaying them as z(y) where z = x-y. I don't know if this follows the conventions of other games; however, it seems it would be less confusing to the vast majority of our players.

I'd say that would be more confusing. When I see 1500 (50) it is easy to understand that you subtract the 50 to get 1450. When I see 1450(50) where 1450 is their real rating, I would just wonder why show the 50 at all?

in your example, the number 1500 has no direct value to players, yet it is the primary number displayed. I suspect that you are biased because you are now used to the TFW interface. however, if this place is to persist, it will need to grow. and most users will be new users.

an analogy: when you ask someone how old they are, do you prefer that they tell you their date of birth?

well, if you wanted it to just be 1450 instead of 1500(50), I would say it makes sense. But putting 1450(50) instead, in my opinion would be more confusing.If the uncertainty is subtracted for you, I see no need to display it at all. I would support it being changed so that uncertainty is not shown and it is just one number, the rating-uncertainty. That would definitely make it easier for new players to understand without having to ask.

Ratings are still misleading. I'd rather see a kill to death ratio against human players. THAT would have some meaning.

For example, as I lurk in the shadows (and on the forum, tee hee) during hiatus periods, my rating drops to nothing because of uncertainty. Victory to loss ratio would be a permanent record.

Yeah I agree with you there Sunyaku. They should at least seperate human wins and ai wins, so that wins actually means something.

A mediocre player beating up on noobs will have a better W/L ratio than an expert playing other experts. The whole point of the Glicko rating system used in TFW (and similar systems, like Elo) is that it takes into account who your opponents are. W/L ratio does not, which renders it meaningless as an indicator of player skill. It also doesn't allow for improvement (because it incorporates your entire history - I don't care if someone took a lot of time to master the game and accumulated lots of losses - I want to know what their level is at the present).

Sunyaku, it seems like your main beef with the current system is that uncertainty is being subtracted when sorting. This makes sense to me - the top players in the list should be the players who are doing well now, not those who have done well in the past, and have not been playing very much recently. However, because the rating is listed as x(y), rather than the suggested z(y), anyone who looks at your stats can immediately see that your position in the list is only due to low activity, (which is a good reason to keep it as x(y)). Basically, the system acknowledges that you play well (you have your high rating, and it appears on your profile), it just doesn't display this as prominently as people with similar, but more recent, achievements.

Well, I don't see the need for win/loss counter, for reasons yaron points out. I would like a separate number for computer wins and human wins. Basically because AI wins are as tough to get as a loss and I like to know how much the person I just played actually wins against other people. Also, I like the x(y) formula basically for the reason yaron said. Let's people know you were once a somebody and you have just fallen from fame. I don't see need for a change there.

I disagree that TFW is truly "glicko". The current rating system is a meaningless indicator of a player's skill because of rating attrition via a perpetually accumulating uncertainty. Put in "glicko" terms, IMHO the Rating Deviation (RD) accumulates much too fast. The proposed z(y) wouldn't be any different. From this perspective, W/L would be better.



Actually, my main beef with the system is the item I just mentioned above: that perpetually accumulating uncertainty exists at all. In chess, for example, a player ONLY has "high uncertainty" until they've played a certain number of games. During this phase, they don't have a "real" rating, they have a "provisional" rating. After that phase, they gain/lose rating points based on the strength of the players they win/lose against. Sure, RD is still a factor, but it should be comparatively low, even if you haven't played in a long time.

I would prefer to see uncertainty removed altogether for anyone who has played 20 games against human players. This would, in effect, make TFW a "glicko" system, as you mentioned. If this style were implemented in TFW, players would only have the second number after their rating if they've played fewer than 20 games. Everyone else would just have an easy to understand number. Then, if a player is inactive for a period of time, they fall off the board due to their inactivity, NOT uncertainty accumulation.

For the most part your post was accurate, but the one thing that is not accounted for is the fact that when an experienced players returns, their REAL rating drops like crazy. For example, once upon a time I was rated over 1700 in both categories. When I came back from hiatus, I had almost max uncertainty. If I WON a game against nearly any player, I would gain very few, if any rating points. BUT, if I LOST a game, I would stand to lose 30-100 rating points a shot. Thus, with a high rating and high uncertainty, it's nearly impossible to drive your uncertainty back down without completely trashing your rating. Since I've been on break again, my current rating and uncertainty level makes me look like a noob.

*Edits: Working as a writer IRL, I can't help but re-read, clarify, and fix typos in my posts. I NEED HELP. (>.>)

As far as I know, the rating system used by TFW is an exact implementation of Prof. Glickman's "Glicko" system, which does feature rating uncertainty (for a description of the system, see: http://www.glicko.net/glicko/glicko.doc/glicko.html ). In chess, they don't have rating uncertainty, because they're using the older "Elo" system.


I'm not sure why you keep referring to rating uncertainty as "attrition". Uncertainty doesn't lower your rating. A 1500(300) player has exactly the same rating as a 1500(50) player - 1500. He will just appear lower on the boards. If he comes back and plays at the same level he did before taking a break, his rating will remain 1500 (just with a lower uncertainty), and he'll move back up on the boards.


That is certainly an argument that could be made. I don't know how fast it accumulates in TFW, so I can't really argue for or against it.
(Jed, if you're reading this, could you tell us the rate you picked for RD accumulation?)


If I understand queeshai correctly, he isn't suggesting changing anything in the actual system, only in the way the results are displayed: display (rating-RD) instead of just (rating).


High uncertainty makes it easier for your rating to move around, in both directions. When an active player wins (or loses) 3 games in a row, the system just assumes that they're having a good (or bad) day, and only slightly shifts their rating. They would need to have several such days for their score to change appreciably. However, when a returning (or sporadic) player has such a streak, the system assumes that the streak (good or bad) is indicative of their current skill, and adjusts the score accordingly. This is because with such a player, the system doesn't have a lot of other recent data to work with.

As for the specific example you cite, note that a 1700 score is extremely hard to hold on to, regardless of uncertainty. I estimate that when playing against reasonably good players (say, 1300-1400), the 1700 would have to win about 9 out of 10 games just to break even (If I have more time later, I'll try to plug some numbers into the equations and verify this). In other words, they would lose 10 times as many points for losing a game, as they would gain for winning a game.

Now, if the 1700 player has a low uncertainty, they might gain 1 point for winning, and lose 10 points for losing. This would mean that a streak of 3 losses might not be so bad (but they'll still need to win 30 games to make it up).

With very high uncertainty, you might look at 7 points for winning, 70 points for losing - which is commensurable with your report. You still need to win 10 games per loss to maintain your score. It's just that the system punishes you more swiftly for not doing so.


I agree that this is a problem. However, I believe the problem lies with the rating assigned to new players. Among established players, 1300 is actually a pretty good score (only achievable by players that have improved their play quite a bit since their noob days). I would say that most beginners play more on par with an 800-900 player.

This means that giving beginners a rating of 1300(350) is unrealistic. It's not very bad, because they have a high uncertainty (so their first losses quickly put their rating in a more realistic range), but it does create some problems:

1. Unrealistic ladder matching for beginners, which is demoralizing for them, and boring for their opponents.
2. The sharp rating drop in the first few games is also demoralizing for the beginner.
3. It makes good but inactive players look bad in comparison (didn't think of that one before reading your post).
4. A beginner that happens to win their first few games (against other beginners), will briefly shoot up into the top 20.

I believe starting beginners with something like a 900(350) score will remedy these problems.