Poll on using AI online

[Chaosrider2808]

The rating system needs to have some kind of external reference, and not just be determined self-referencially.

In person, the self-referential scheme works fine. On line it's very different, and works much less well, and must be supplemented. As a crude low level example, the Igowin rankings don't change, although there's probably a better standard.

As I write this, I realize that I don't think I understand the problem that you're concerned about. Could you elaborate a bit?

Thx,

TCS

[Uberdude]

External anchors for a rating system are not what I am talking about (though I don't see how a self-referential online system like a go server is any worse than a self-referential offline system like EGF or AGA grades).

Let's say we have Bob, a human 5k on KGS. On a good day maybe he is 4 or 3k strength and can occasionally beat such players on even. On a bad day maybe 6 or 7k. When he's drunk maybe 8k (or maybe 4k, I know someone whose drunk account was higher ranked than his regular). But 5k is a sensible average so if another 5k like Mary wants to play a reasonably well matched game online then the ranking system pairs them and they can enjoy a nice game.

Now let's say Bob on his 5k KGS account uses LeelaZero (super pro level) 1 game in 5. If he is paired against 5k Mary he absolutely smashes her and she has no fun. Next he (not LZ) plays and loses to another 5k, then beats one. His rank will go up a bit from the LZ wins but not much seeing as 4 games in 5 he plays like the 5k he is. I think this is a bad situation and should be avoided, a primary purpose of the ranking system is to make well-matched games and by sometimes using LZ Bob is perverting it. Also if his rank went up to 1k* then when he plays real 1ks even they have a less enjoyable mismatched game against a 5k (or get smashed by LZ).

*If the LZ wins are only against ~5ks then I doubt it will get this high, at least not for a long time, so not the 1/5 weighted average of 5k and 15d. If on the other hand Bob somehow persuaded 9ds to play him with his 5k account even and then beat them with LZ then his rank would go up a lot. But if he sometimes loses to 5ks then you are basically asking the ranking system to answer an impossible question to which there is no good answer.

[gowan]

There seems to be widespread objection to the practice of "sandbagging". So isn't playing with undisclosed AI assistance effectively sandbagging?

[Uberdude]

Sandbagging is (intentionally) playing with a rank lower than your strength, usually because it's fun to beat up weaker people (and to do this you resign games you are winning). There are grey areas, such as temporary rank inaccuracies, an inaccurate low rank because you timeout a lot, a server not allowing you to register with an accurate rank so you have to work your way up through the ranks (e.g. Haylee 3p starting as a 5d on Tygem because that's the highest you can register, I don't count that as bad sandbagging), or playing on a low-ranked account for teaching (e.g. dwyrin, I think this is kinda fishy sandbaggery if you don't tell the opponent, do the ends justify the means?). So if your account says 5k and you are sometimes 15d LeelaZero then yes that's sandbagging. But if you are always LeelaZero (disclosed or not) and your account says 15d that's not sandbagging, it's just an accurate rank. Or if you use LeelaZero to make your account 15d and then start playing with your 5k human moves that's what I call heliumbagging.

[Chaosrider2808]

Ah, OK, so you're talking about an automated matching system that matches people based on the rank that it has for them. That makes more sense. That was probably clear from the early discussion , but I didn't pick up on it.

All automated systems have a higher "risk" in this regard than in-person playing, particularly any group of people that you play with regularly. There are social clues that you see in-person that we can't yet transmit online. Social pressures as well.

So, the "payoff" for diddling with your ranking is that you get to beat up on lesser players?? Really?

How prevalent is that disorder?

Thanks for the clarification!

TCS

[Uberdude]

Ah, OK, so you're talking about an automated matching system that matches people based on the rank that it has for them. That makes more sense.

Not particuarly an automated one, manual ones are affected too. An open games/challenges list like on KGS/OGS lists a game offer from Bob [5k]. Mary is 5k and wants a game. She clicks on Bob's offer and challenges him because she wants to play a 5k. She plays a 15d LeelaZero instead and is sad.

[Chaosrider2808]

My apologies; I used sloppy language.

What I actually meant to say with the word "automated" was "not physically sitting across from each other." That's the important distinction. Any communication that isn't face-to-face is subject to an entire range of IT security risks.

TCS

[jlt]

What can someone gain by using the AI during the game and not after it?

I have never used an AI to play online, but I suspect that most people who use an AI do it for one or several of the following reasons:

• 1. to boost up their rank in order to play stronger opponents
  2. in order to brag about their rank
  3. to alleviate the frustration of losing.

Another way to reach these goals is to play many ranked games against bots: since bots tend to repeat the same mistakes, once you find a way to trick them you end up winning more games than you should given your rank.

Sandbagging is another way to reach goal #3 (however it is better to play against people who volunteer in the first place to play against much stronger players).

[Uberdude]

What I actually meant to say with the word "automated" was "not physically sitting across from each other."

This entire thread is about "using AI online", the clue is in the title

[Chaosrider2808]

Yes. I was making an observation that was tangential to that topic, but not directly on point.

Sorry.



TCS

[gowan]

Sandbagging is (intentionally) playing with a rank lower than your strength, usually because it's fun to beat up weaker people (and to do this you resign games you are winning). There are grey areas, such as temporary rank inaccuracies, an inaccurate low rank because you timeout a lot, a server not allowing you to register with an accurate rank so you have to work your way up through the ranks (e.g. Haylee 3p starting as a 5d on Tygem because that's the highest you can register, I don't count that as bad sandbagging), or playing on a low-ranked account for teaching (e.g. dwyrin, I think this is kinda fishy sandbaggery if you don't tell the opponent, do the ends justify the means?). So if your account says 5k and you are sometimes 15d LeelaZero then yes that's sandbagging. But if you are always LeelaZero (disclosed or not) and your account says 15d that's not sandbagging, it's just an accurate rank. Or if you use LeelaZero to make your account 15d and then start playing with your 5k human moves that's what I call heliumbagging.

I guess I was thinking of the case where I declare a 5k rank and use LeelaZero in free, unrated games. That would be sandbagging.

[EdLee]

Some folks have an exactly opposite definition ?

[Bill Spight]

Some folks have an exactly opposite definition ?

I'm pretty sure the Urban Dictionary is wrong, written by people who do not know the practice first hand. Sandbagging works this way. First there is deception, either acting as though you do not play as well as you do, or acting as though (in poker, bridge, and other card games of imperfect information) that you do not hold as good a hand as you actually do, and then later metaphorically dropping a sandbag on your opponent or opponents by revealing your strength. In go the deception comes before the game, in poker or bridge it typically comes during the play.

[EdLee]

Hi Bill,

6 & 7 here.

[gowan]

Some folks have an exactly opposite definition ?

I think most complaining about sandbagging in go comes from double digit kyu players who encounter an opponent rated double digit but playing much stronger. Sandbagging comes from someone with sadistic intent who likes making weak players suffer. It was mentioned previously that the mechanism of keeping the sandbagger's rating down is by deliberately losing games. The urban dictionary used the word hustler in this context but I think that usually applies when money is at stake. A skillful hustler lures a mark into playing a game, say of pool, by deliberately playing badly losing money, perhaps with a collaborator. The mark is offered a game at money stakes which the hustler loses, then asking for another game at higher stakes which the hustler wins. Part of the skill of the hustle is to get the mark to think that the hustler's win was a fluke and the mark is lured into a game with even higher stakes.