Zen19 had made 3d on kgs!

[CarlJung]

I wonder whether they'd be more fun to play against if there were a "play to maximize expected points" option (instead of "play to maximize expected winning probability"). Should that be easy for developers to implement?

That's what the previous generation of bots did and it only got them so far. The monte carlo algorithm changed that by focusing on winning probability and suddenly beat all the old bots. It's not possible to change MC to maximize the win (and perform better than last generation).

Btw, what would be the biggest win? Killing all enemy stones of course. Try that yourself in your next 10 games and watch your rank plummet

[prokofiev]

I wonder whether they'd be more fun to play against if there were a "play to maximize expected points" option (instead of "play to maximize expected winning probability"). Should that be easy for developers to implement?

That's what the previous generation of bots did and it only got them so far. The monte carlo algorithm changed that by focusing on winning probability and suddenly beat all the old bots. It's not possible to change MC to maximize the win (and perform better than last generation).

Btw, what would be the biggest win? Killing all enemy stones of course. Try that yourself in your next 10 games and watch your rank plummet

Thanks, this is interesting. Maybe there's some intermediate though?

For example: What if you use the current "winning probability" method, but each turn, adjust komi to the value such that you think your current winning probability is roughly 50%, and then make the move that has the highest winning probability.

Variants:

-Only let the komi float in one direction (e.g. still play stodgily when you're ahead as current bots do, but let the komi float if you're behind).

-Put a cap on how far you'll let the komi float. (Maybe different caps in the two directions, generalizing the previous point.)

Does this do anything?

Edit: More variants:

- You probably don't want to lose the edge computers have at the endgame (over similarly ranked humans), so e.g. let komi float less as you get to the endgame, or only start the komi floating if your expected win percentage gets rather low (i.e. the range where computers do strange things).

[Mike Novack]

I got a registration key for MFOG 12 as a prize for winning a local tournament. I've only played a handful of games with it, time settings 15 minutes, 5/30. In the one even game, at first it seemed to be a reasonable facsimile of AGA 2-3 kyu, but somewhere in the middle game (neither side having a significant advantage in my view), it started playing bizarre responses to my moves, and completely collapsed. The other games I gave it 9 stones, just to see if I could win (I couldn't), and 3 stones, where I might have won if I hadn't been playing late at night and just messing around.

Lindentree --- What version (update) are you running? And what are you running the program on? (machine power). If the version of MFOG12 you downloaded was the original may need some of the fixes. I think the Feb 14th build is the latest version (you don't need to redo the key to do this uninstall/reinstall). Also machine power is critically connected to the time setting for the machine if machine power isn't above some threshhold. That's probably why you are allowed to set seperate time controls for the players. For the program to play as strongly on this old desktop (far weaker than "standard" for this program) as on one of my laptops (which are "standard" for the program) I'd have to quadruple its time allowance.

[lindentree]

Lindentree --- What version (update) are you running? And what are you running the program on? (machine power). If the version of MFOG12 you downloaded was the original may need some of the fixes. I think the Feb 14th build is the latest version (you don't need to redo the key to do this uninstall/reinstall). Also machine power is critically connected to the time setting for the machine if machine power isn't above some threshhold. That's probably why you are allowed to set seperate time controls for the players. For the program to play as strongly on this old desktop (far weaker than "standard" for this program) as on one of my laptops (which are "standard" for the program) I'd have to quadruple its time allowance.

I have the latest version, 12.020, running it on a laptop with a dual-core 2.0 Ghz processor. I probably should test it on different time settings/handicap, but honestly I have better things to do. I'm glad I didn't pay for it; the extra features aren't worth it, although I'm going through the (substandard) problem collection as a supplement to my daily L&D regimen.

[yoyoma]

Zen19N (the one that plays 30+5x:30) is on KGS now, played and won 2 games as a 1d so far.

ETA: Looks like it is playing 20+5x:30 now. Still very reasonable time controls for online games.

[Mike Novack]

Try a time setting of 30 minutes, 5/30 and see what happens. Or 60 minutes (for all).

This sort of algorithm is very non-linear with time (because the reliability of statistics relative to sample size is very non-linear). Your computer power is "standard" for the program and you have the current version so those aren't the problem. But most of these go playing programs look at how much time they are being allowed and adjust their sample size accordingly. Not quite enough time and performance will deteriorate rapidly (but above a certain threshhold improvement with more time slight).

You should be seeing better than ~2-3k AGA. I'm running MFOG12 on a comparable machine and I certainly would be noticing if not doing significantly better that that (I need the same handicap against MFOG 12 that I do against a human 1d AGA).

[Javaness]

Back in the chess world, huge advances were later made, and of course Kasparov lost a famous match. But Garry complained bitterly at the time and ever after that the Deep Blue team refused to let him see examples of its previous games. The assumption is that the team knew that they could be beaten once its quirks were known.

I expect the same pattern to apply to computer go. I don't accept the MoGo and Zen ratings except as short-term indicators (and, yes, impressive ones at that). Once they go commercial I expect their grades to drift down far and swiftly, just like my first copy of MFOG.
...
We may get a clearer view in London this Christmas. There is to be a ?ten-game match between John Tromp (a 1 or 2-dan, I think) and the best computer program as chosen by a computer person - for money! I think John Tromp's money is safe, although he may get bitten in the bum in the first game or two if he hasn't seen that program before.

Kasparov played very badly in his match against Deeper Blue, but he, or any other world champion, would certainly be annihilated in a match now against the top computer chess programs in the world..At the time, I don't think IBM chose not to agree to a rematch because they were afraid of losing, it was more likely that they had nothing to gain by a rematch.

I don't think UCT based engines are as easy to exploit as the old generation of computer programs. Zen is a commercial available program already, I haven't see anyone on KGS systematically trashing it. I would back the computer over Tromp myself. Since I won't be able to watch the match properly, I doubt I will have a bet on it though.

[pookpooi]

Speed game is not computer advantage, it's human fault.
Computer can play 1sec/move or less while human can't.
If you think speed game is not fair to human then give computer infinite time.
Of course those matches are not 'enjoyable' by your taste but in computer there's no 'enjoyable' or 'unenjoyable' time.

[Mike Novack]

ROFLOL --- Well not infinite time. The whole point of the problem not being exactly solvable is that the solution has to be within some finite amount of time.

I think we need to pay some attention to the particular time settings that the developers of this sort of software choose to have their bots play at. They probably know where, given the speed of the machine used, the points of diminishing returns lie. Also of course related to the willingness of human users to agree to playing at those time controls.

The machines used make a big difference which you can roughly translate into the time the machine is allowed (only roughly since there is overhead combining the results from multiple processors). On one hand I do have interest on how well these programs can perform on massive machines but more interest in their practical strength (running on a machine of reasonable power that an end user could be expected to possess).

Thus if the overhead is ~20%, the the program running on a thousand core supercomputer at 10 seconds/move might equate to the same program running on a 2 core machine at an hour per move. Which should be an indication how severe the diminishing returns are (how small the benefit from additional time above some critical amount).