stuck at 13 kyu Many faces of GO

[1986]

its been a while since i played MFOG i decided to do a rematch after i got offline yesterday, i pretty happy i won i never beat this level yet until last night, but i have been working hard at my go doing tsumego, online games, and elementary books i felt stronger since i played last and i expected to win this game. i should have pushed harder i wanted to win by more points but i didn't see any profitable attacks so i played it cool. also in the middle game i didn't play aggressive enough so there were no good invasions or cutting opportunities.



well im not stuck anymore so thats good news
now i want to make it resign next time!

[Mike Novack]

I am going to suggest something (in my opinion how to best use a program like this)

a) Use the computer program weakened to about your own strength to play against even only as a substitute for not having a human opponent available. Otherwise play against humans when playing even.

b) To maximize learning set the computer program to be several stones stronger than you and take the appropriate handicap. This will force you to learn to use strones already on the board, maintain connectivity, make territory by attacking, etc. and when you make a mistake the computer will likely spot it and punish you.

Used this way the computer won't be playing as differently than a human would as in "a" so less likely for you to learn bad habits.

c) As soon as you get strong enough you will want to be playing MFOG when it using "Monte Carlo". If you have updated to MFOG 12.021 the top two levels (1 dan and 3 kyu) are doing that so you are getting close. You could try the 3 kyu setting,take 9 stones, and see what happens.

[Li Kao]

Does ManyFaces have some kind of dynamic komi? Because from what I heard handicap doesn't work well with monte carlo AIs.

[Numsgil]

Does ManyFaces have some kind of dynamic komi? Because from what I heard handicap doesn't work well with monte carlo AIs.

Er, that's weird if that's true. Unless you mean for hybrid monte-carlo where it can't use an opening move dictionary?

[Mike Novack]

Not sure what is being said.

a) I am not at all sure how Fotland has managed to get two levels of play where the selection algorithm is a Monte Carlo tree search. But I can think of a couple possibilities beyond "dynamic komi".

b) MFOG uses an AI engine to generate a plausible move set and then (if at a level using Monet Carlo) using Monte Carlo tree search to select the best of the lot. So yes I can (and probably does) use fuseki data bases, etc. Very much "hybrid" with the intent that the play be more human like.

[Cassandra]

Does ManyFaces have some kind of dynamic komi? Because from what I heard handicap doesn't work well with monte carlo AIs.

Yes, it is said to be "linear" Komi.

E. g. with 8 stones handicap Many Faces starts with a Komi of 50.5 and this Komi is reduced by 1 point every two moves. With move 100, Komi remains 0.5 points. Reduction does not depend on the situation on the board, so "linear".

May be that there are several parameters that can be set (start level, decrement, ...).

[Mike Novack]

Separate issues? (and I really think discussion at this level belongs in the "computer go" section)

1) How a Monte Carlo tree search based program works at all (or how best works) might depend on using an "evaluation komi" different than the actual komi of the game. It would be an empirical matter whether slight skewing of the outcomes of the random games helped or hurt.

2) That's why I ignore this in discussing what sorts of adjustments might be needed when playing when handicaps are involved. There are all sorts of possibilities with "dynamic komi" besides starting with some set value other than the actual komi of the game and then a deterministic adjustment back to the actual. As I indicated in "1", I don't know if one would ever use the actual. I'm not looking at the code and then deciding what the algorithm is!

3) I was discussing another issue, how to have a Monte Carlo tree search based program that had more than one playing strength level. That's nowhere near as simple an issue as it might at first seem

[daniel_the_smith]

3) I was discussing another issue, how to have a Monte Carlo tree search based program that had more than one playing strength level. That's nowhere near as simple an issue as it might at first seem

Is there something wrong with just using different amounts of time for the different levels?

[Mike Novack]

3) I was discussing another issue, how to have a Monte Carlo tree search based program that had more than one playing strength level. That's nowhere near as simple an issue as it might at first seem

Is there something wrong with just using different amounts of time for the different levels?

Yes -- unless all you want is a very small adjustment. The performance of the algorithms used is very non-linear with respsect to time. Take the newest release 12.021. That will max out at about 2 dan but Fotland says that's on a 12 core xeon. It is supposed to play at about 1 dan on a "standard machine" (2 core, 2+ GHz). In other words, giving it (in effect) 6-8 times less time). But 12.021 is supposed to be offering a 3 kyu (as well as 1 dan) level.

Note that there is also a problem with the (lower) AI levels too. If one takes the simplest approach of leaving something out of consideration (that the higher level has coding for) the playing behavior becomes less human like --- one can learn how to beat the program and the program can't "see" how to learn to avoid this (it can't).

I'm not "in the game" of writing these things but if I were the first thing I would try would be to introduce "probability of making an error in play". More like the behavior of a weaker human who sometimes sees the best move and sometimes doesn't. So at the AI levels, at move N it might or might not be using some parts of the engine. At the Monte Carlo levels, a probability that it might choose to make the 2nd best move instead of the best. I don't know that this approach would work but if it did would surely be very easy to adjust.

This is my last comment here (this discussion needs to be moved to "computer go")

[daniel_the_smith]

3) I was discussing another issue, how to have a Monte Carlo tree search based program that had more than one playing strength level. That's nowhere near as simple an issue as it might at first seem

Is there something wrong with just using different amounts of time for the different levels?

Yes -- unless all you want is a very small adjustment. The performance of the algorithms used is very non-linear with respsect to time. Take the newest release 12.021. That will max out at about 2 dan but Fotland says that's on a 12 core xeon. It is supposed to play at about 1 dan on a "standard machine" (2 core, 2+ GHz). In other words, giving it (in effect) 6-8 times less time). But 12.021 is supposed to be offering a 3 kyu (as well as 1 dan) level.

I think if you continue to decrease the time (or, more precisely, the # of playouts) the performance will degrade quite a bit, eventually to the point where it's not much better than random (one playout per available move). Non-linearity doesn't matter if you're trying to make it play *worse*... (this may not apply completely to MFoG as I don't think it is pure monte carlo.)

Whether those mistakes would be human-like or not I can't say, monte-carlo doesn't play very human-like to begin with (though I haven't tried MFoG, maybe it's better than the others).

[1986]

thanks for the info i didn't even know about the update so i got it, but the updated version still plays like the old version to me the dan level must have been the only upgrade 12 kyu still plays the same its about the same amount of aggression and style., but you say time and the speed of the computer is a factor huh, so the game will play better if i add more time to it, interesting... i think the next time i will try to play a long game with the dan engine with a challenging handicap and see how i do thanks everybody for the comments & tips

[Mike Novack]

I said I wasn't going to comment more but need to correct a misconception.

When I said that time/computer power was a factor I meant at the levels where Monte Carlo tree search is involved. I do not mean at the lower levels of MFOG where it's just the AI "go knowledge" engine and positon evaluator choosing the moves. Nor do I know what adjustments Fotland may or may not have made to the lower levels in terms of what they were set at. Nor do I know that the AI "go knowledge" engine hasn't been changed -- an improvement there might not change the level of agrression as that could be up to the AI "position evaluator" component which does the actual move selection.

If you have been playing even against the 12 kyu level but now have update 12.021 and what to see what the "Monte Carlo" levels are like (should be much less passive) I suggest you try the 3 kyu level and take 9 stones rather than the 1 dan level and even more stones. Let us know what happens.

I will add what I am experiencing. When playing against the 1 dan level it appears the "right handicap" is more critical than I would expect against a human opponent. In other words, I win rarely at one handicap and almost all the time at one stone more.