http://www.unbalance.co.jp/igo/sigoMeijin/

Does anybody know more about this? Specifically, what kinds of engine improvements went into this and what win rate 2015 manages versus 2013 on the same hardware?

Hope there'll be an English release soon.

They're saying また、どなたでも正しい手合いでの対局が行えるよう、14級～六段の20段階のレベルを用意しました。
Which I think means: 'Furthermore, in order for it (Crazy Stone) to be able to play adequately against anybody, we organized it into 20 levels, from 14kyu to 6 dan.'
Obviously, this 'information' has to be treated with caution since:
_ this is an advertisement
_ we are talking about japanese ranks
_ perhaps the ranks don't refer to Crazy Stones' possible levels but to the human's who will be playing with or against handicap stones
_ Crazy Stone is just a software which strength obviously depends on the user's hardware

Anyway, I didn't see any clue on to how much stronger this software is compared to the previous version but they say it has better endgame, which could mean that it won't play stupid paranoid moves inside its own territory anymore and won't win by only half a point when it could have won easily and safely by a much larger amount.
Which also means that when losing by a few points only, we'll be able to say 'Phew! that was close' and be right about that.

_________________
Bored with skillful play? Watch my blunders on my YouTube Channel

The advertisement is most certainly not taking hardware into account.

When CS is playing at 6 dan in tournaments and on the server, the bot is running on hardware far more powerful than home users are going to have available. In a case like this where the program is computationally intensive, that makes a big difference. On the other hand, "diminishing returns" on crunch power has already been reached.

So while it would be less strong playing on a home user's machine, I'd guess about 4 dan on a powerful home machine, maybe 3 dan on an ordinary home machine. On the other hand, if you home machine is borderline workstation, say a cpu like an i7-4790 or better, possibly 5 dan.

What happens if you run CS on a supercomputer like this http://www.top500.org/system/177999

You find the Divine Move.

... or because CS is probably not written to support parallel computing it may just run regular speed on one processing node (which may be less powerful than a $2k home desktop computer).

_________________
I am John. John-I-Am.

I was hoping for that

But what if in theory CS was changed so that it can do parallel processing?

Would monte carlo tree search benefit from 10^15 calculations per second?

Possible misunderstanding about the differences between the theoretical algorithm and its practical (computationally limited) modifications.

a) The theoretical algorithm, using random play between two equally bad (random) opponents would find the "divine move". But the number of playouts AND the breath of the tree enormous. Even super computers not fast enough.

b) Practical implementation use pruning and favoring adding to the tree in certain directions (say down a ladder to determine that) with these "tricks" adjusted to the time limitation. In other words, specific to the amount of processor crunch power on hand. The same program (the same set of tricks) might be useful over a reasonable range of computer power, but wrong choices if orders of magnitude different.

Understand? Increasing the computer power for the existing CS by an order of magnitude might slightly increase it playing strength by decreasing the (already small) percentage of time the wrong move chosen because too few playouts. But it won't help the situation where the "tricks" have accidentally pruned away "the divine move".

BTW, all of the MCTS programs out there are capable of utilizing multi-threaded processing. With almost all the crunch concentrated in doing a playout and additions to the tree possibly asynchronous, the algorithm is very suited to parallel processing. Before MCTS came along, go++7 was perhaps the strongest of the go AI's. The reason why there was no MCTS go++8 released is that Mick Reiss didn't have the experience to get the multi-threading right (so it was too slow; I was beta testing for him)

Crazy Stone used 64 cores to play Ishida Yoshio in 2013. That doesn't mean it would be ready to scale well on to something like the Tianhe-2, but it has at least shared-memory parallelism.

https://gogameguru.com/crazy-stone-comp ... -4-stones/

But I think the problem would scale well, precisely because the computation of doing a playout is massive compared to recording the result of that playout and adding its first move to the tree (one tree managing process could use a large number of play it out "slaves". And the problem is computationally intensive, but not one requiring much memory*.

And if you wanted a reasonably powerful machine (on which CS would play ~5 dan) it might not even cost you as much a $2K if you knew where to look for reconditioned "workstations" and had some patience till a good one showed up at a lowish price (in other words, I couldn't get one today perhaps, but very likely within a couple months)

* It's not when under way with MCTS that these programs are using much memory. They are starting out with a fair sized fuseki "book" and I bet that's when memory is being used (to save on time lost doing I/O)

So apparently there are 3 improvements over the previous version:

- Updated version of Crazy Stone (hard to say how much of a difference)
- "Tesuji study" which is apparently an improved analysis mode
- "Dojo mode", which disables undos and hints and keeps track of your win/loss record

Not a whole lot, but I might still get this version in Japanese. I regretted going for the English translation of the previous one, as they completely removed the Japanese byo-yomi voice and did not replace it with anything else which made playing blitz games impractical.

Oh, and BTW, don't simply count cores. Try to look up "crunch benchmarks" for the machines being compared. Most of the machines being used by the top programs in these competitions aren't all that powerful.

It wouldn't be all that expensive to get a machine with 1/2-1/4 the power, so playing just a stone or so weaker. The reason for that is you only need the crunch of a powerful cpu, not all the other expensive components a server might have, like a vast array of hot swap drives, etc.

Just curious, will you never be happy until you get exact cpu model number, bios, current running temperature, operating system build rev, and quantum effects in the exact room being used?

You reply to all of these messages in the same way. Most of us can use a core count and year to get a rough idea of horsepower. You seem to need down to the hertz level for some odd reason.

No, I was referring to crunch benchmarks, not the details. For the simple reason that just "core count" does not give a good estimate of that. Hertz (clock speed) doesn't give a good estimate either, not over very different architectures.

In spite of the fact that some of the strongest machines being used to run these programs have 32-64 cores those are relatively weak cores individually so the total machine crunch power is at most a few times, not order of magnitude times the power of the most powerful 4 core (8 logical) cpus. That sort of thinking is important for those of us considering whether we could afford a machine strong enough to run a bot playing at several dan strength. In other words, from my point of view the question is how much more powerful are these machines than the most powerful affordable machines you and I could have if we needed them.

Look, there are some sorts of problems (computer problems) where a large number of cores is important because the overall problem has a great number of different processes in parallel. But in this problem the crunch components are the playouts and those are the same process in parallel. Which makes the speed/power of the individual cores and the number of them a direct trade off.

PS --- core count and year ignores "purpose" of the cpu, what other requirements is the design trying to satisfy. For example, cpus for our portable devices have to consider power consumption and heat to be dissipated. So no or minimal increase in power but half the power needed to achieve that (so twice the battery life) could be a great advance.

Thank you for answering me! The answer is yes, no matter what answer you get, you will respond the same way.

I think the Japanese version would be too bothersome for me, but this might be the first version I'm tempted to shell out the 80-something bucks for.

What a fool I am! I was originally given a Japanese copy of the pre-2013 (2011?) version as a present. I later bought 2013 in English for the upgrade and presumed ease of use. I have used it mainly for analysis but recently when I tried to play against it one evening, I thought that I had screwed up the configuration somehow so that it no longer counted boy-yomi. I was sure that I remembered it having that feature in the past. Now you tell me that I paid $80 to turn off the voice. LOL, shame on me!

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

I bought it. I'm getting crushed with 9 stones again. I hate this game

One thing that is still annoying: the byo-yomi voice is only active if you activate the voice announcement of every move too, that's a lot of talking.

If he follows chess engines at all, I can see where he's coming from. In the Chess engine sub-culture, the differences between engine strengths at the highest level might come down to 20 to 40 ELO, at a level in the 3400+ range.
For comparison, a "C" rated chess player is 1500, the Top 10 players in the world are around 2750 to 2850, the chess engines are around 3300 to 3400 ELO. So 20 or 30 ELO is statistically insignificant to a human, but apparently means a LOT to those people that are interested in chess engines. Even then, your talking about chess games where there might be an 50% or higher draw rate. Yet if an engine running on some nice rig might win an extra 1 or 2 games in a match of 100 games, those people that follow that stuff celebrate the accomplishment.

Chess is more complicated then GO, as far as the rules and piece movement, but has a fraction of the *derivatives compared to GO, so some things have been vetted out of chess engines. For example, using **Monte Carlo to run thousands of games in order to hopefully find on accident a best move doesn't work. That's also why scaling engines to use more and more cores doesn't really work. Although using more cores does work to a point. Using 64 cores compared to 32 cores gives you statistically zero gain. Because chess is an INT operation, you can't really use parallel processing, so using GPU is more or less worthless, since they don't do INT operations. Throwing a chess engine on Super Computer is all but worthless. They've even done distributed chess games which used tens of thousands of computers across the internet. Although the games were largely successful, it showed that the total increase of the distributed engine's power was fairly insignificant compared to the total resources used. -And a few years later a modern chess engines would handily beat that particular program.

*I think I read someplace a chess game has 10^50 positions and GO has 2.81*(10^170)

**Monte Carlo has been used for some effect in post game analysis, in which the analyzing program has unlimited time to compute and compare Monte Carlo lines to the actual play. But in those cases, the program isn't looking for a winning line, but merely a better reply to what was played in the actual game. If a outright mate or forced draw isn't obtained, eventually a game of chess will lose pieces until so few are left that an endgame table base can be used for perfect play and instantly the chess engine can tell you if its a win/lose/draw, and spit out the optimum move sequence.