This Topic is dedicated to all of us how still lose constently against computers that are supposed to be weak at go.

Lose constently... till there is that game that you won and you are so.... happy
So here is the one for me

Ok ... strong people do that all the time but come ! let's beat those computers !

To be honnest I still used that "estimate score" function from time to time so I still need to beat him without any help... but I'm still happy...

_________________
Converting the book Shape UP! by Charles Matthews/Seong-June Kim
to the gobook format. last updated april 2015 - Index of shapes, p.211 / 216

Congratulations! One reason that I quit playing correspondence chess was the introduction of personal computer programs that play at a grandmaster (or higher!) level. I suspect, however, that at my age I will never catch up with the PC Go strength curve. So, battle those demons for all of us doomed to be patzers forever!

Is that the computer program playing as strong as it can? Or have you weakened it (used a low strength setting) so you could play it in an even game?

Just my opinion, but I don't think that's how to make best use of a computer program as opponent. The problem is that you are both making mistakes at about the same level. It's one thing to be playing a human opponent an even game for the thrill of making fewer mistakes than the other guy and so winning. But ideally you would have a human opponent several stones stronger than you are, no longer making the mistakes typical of your level, so you can learn the consequences of making those mistakes. Of course you would need a few handicap stones to make a game out of it, but that's not all to the bad. With 3-4 handicap stones you would still be learning to apply joseki and with high handicap learning how to keep connected.

Set your computer program somewhat higher, enough so that with 5-3 stones ypou sometimes win, and when the wins become 50%, drop a handicap stone. See if that doesn't help you learn how not to make the same mistakes over and over again << make a mistake but the opponent doesn't punish you does not teach you to stop making that mistake >>

It's at its strongest level (the programm is "Igowin pro" running an Iphone 4. the level is annonced to be 3kyu but I sometimes play a guy that is 4kyu in my club and I still need 9 stones...



Most of the time, I still lose against that computer... The problem is that I have not that often time to play a real person... only one game per week against a real human for me... but I have plenty of time in the train every day, I mainly read books and play tsumego, and sometimes I play a quick game like this one.



100% agree with that



I also own "smartgo kifu", but I used to find it weaker then Igowin... but I think I can change its level in the settings... I will have a look at that, as I prefer the computer to be the strongest possible and play with hanicap stones.

_________________
Converting the book Shape UP! by Charles Matthews/Seong-June Kim
to the gobook format. last updated april 2015 - Index of shapes, p.211 / 216

It uses a challenge system, so to get a smarter AI you have to win the weaker one first (Or I just missed the option to set ranking. I am hardly winning on the lower levels though).

Computers are of 'course not as good as a human player. But we can learn something about ourself. Oh and counting scores and assessing where we are in a game. (Counting points is one of the harder parts of the game to actually understand in my eyes. I am still not sure I exactly understand it)

I think you need to check more thoroughly. While I am not familiar with Igowin or any of the other programs intended for weak* computers I do use the related program MFOG on a real computer, one with enough power that it's strongest level is about 1 dan KGS. Fotland would not be bothering with Igowin if its strongest level were making as many serious mistakes as this game indicates.

One thing I learned early on in my days in the cypher mines, when the results are impossible, one of the things you are assuming to be true isn't.

Things that might affect playing level (things for you to recheck)
a) Reread the instructions for setting playing strength and how that is reported to you.
b) Check the time settings (how much time are you allowing the computer for its moves). Unlike some of the other programs where "level" determines number of playouts and hence time, I believe that Fotland has his programs take the time allowed as a given and from that determines how many playouts can be used. So the program, if using MCTS plays weaker (in an absolute sense) with less time, and the point here is that MCTS collapses more or less completely if not enough playouts and I believe that Igowin pro does use MCTS for its strongest level. In other words, you might try a setting a couple levels weaker (those would be AI, not MCTS) and if you find that the program plays better, that's your problem.
c) Related to that, is your iphone doing anything else while you are playing against Igowin? Again, I am not at all familiar with these small machines so don't know what the possibilities might be. But if concurrently busy with other tasks, that might be the source of the problem.

In other words, the 3 kyu advertised strength would be for when given a normal amount of time (and that the same as you have) and with the machine not being asked to do anything else concurrently.




* Sorry , but your iphone is basically a powerful communications device, not enough crunch power to be fully effective using MCTS.

Mike, Igowin in theory handles the time taken with respect to level, and autoadjusts its level, so it is probably depth-based/number of simulations based and not based on time. This can be clearly seen from Igowin in my old iPad vs my iPad Air.

iOS' multitasking is "weak multitasking." So, essentially, if you are playing go most of the resources are devoted to that. There are cases when this doesn't happen, but they are somehow corner cases (an app can request up to 10 minutes of "background time", also, music/audio keeps playing in the background, as well as audio recording)

_________________
Geek of all trades, master of none: the motto for my blog mostlymaths.net

I tryied again but I really cannot change the level to be stronger then 3kyu , but I saw that I had a komi of 0.5... maybe that change the way igowin plays...
I also just played a game on IGS vs a guy that is 2kyu, he gave me 4 stones and I ... did my worst game ever... enjoy...

that all started at move 13 that I excpected to be on E3 or F2 and so I would have play G5 like in the books... but that went... another way... F4.
I suppose 14 on F5 is ok as I cannot be cut, then 15 at G5... and then good bye...

really really nothing to compare with Igowin...

[edit]
My best move is 112
[/edit]






I really enjoy that kind of game, that give me way much more material to study then what the computer gives me !

_________________
Converting the book Shape UP! by Charles Matthews/Seong-June Kim
to the gobook format. last updated april 2015 - Index of shapes, p.211 / 216

OK, you have some experience with this program.

Was your experience that even at its highest strength setting it was far too weak to be of any use for you*? Look at the game Oca first presented. Based on your experience playing against this program does that look to you to be Igowin playing at its strongest level?

If not, then isn't something going wrong for Oca? (about getting Igowin to play at its strongest level)

Oca, is there any sort of "help" forum for Igowin? Your question would be along the lines of "mysteriously playing much weaker than advertised" <<if, as I expect, other Igowin users will chime in to confirm that the game you first posted can't be Igowin actually playing at its strongest level>>


* You are much stronger than Oca.

The version I have checked is Igowin HD, which is supposedly the iPad version of Igowin (Pro is for iPhone.) They are supposedly the same engine, but just seeing the first few moves I can tell it's very different. Just yesterday I played a game after supper with its maximum setting (which is indeed 3k) and lost by 5 points (I was black, komi 6.5.) I didn't give much effort to it, but the game was decent on the computer part. It showed quite clearly to be a MCTS: I made a reckless invasion that died, at which point I was ~20 points down but had sente, so I could reduce its lead... and of course, MCTS kicked in and he let me reduce as much as I liked as long as it won anyway.

Igowin didn't make any really stupid/odd moves during fuseki, unlike what can be seen in oca's game, where move 6 is already very weird (not counting the weird choice of opening the computer did, though.) I don't know why it did in that case, but as advertised, the strongest level is 3k. No way to manually set it any higher, although David's page advertises its strongest level as 1k. Weird. A decent guess would be that the iPhone version has a smaller fuseki pattern library.

Utility of playing against it? None. I only use computer play for 9x9, where its level is more or less on par with very good players. Since finding good opponents in 9x9 is not that easy, a quick, computer game of 9x9 is most enjoyable. Probably also decent at 13x13, but playing 19x19 is pointless. As far as computer opponents, I much prefer Crazy Stone for iOS. Its level of play is more consistent, although when it blunders, it blunders big.

_________________
Geek of all trades, master of none: the motto for my blog mostlymaths.net

Here is another game where Igowin won.


is different in this one, but I don't know if it's good or not...
at what I remember, Igowin never approched my sanrensei with

Click Here To Show Diagram Code

[go]$$W
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . 1 . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

_________________
Converting the book Shape UP! by Charles Matthews/Seong-June Kim
to the gobook format. last updated april 2015 - Index of shapes, p.211 / 216

Well, it's unconventional. Hard to say bad or good (it's not "specially" good, as in not usual with pros) but not as weird as previous 6

_________________
Geek of all trades, master of none: the motto for my blog mostlymaths.net

Crazy Stone running on what hardware.

We cannot properly compare these programs without reference to the crunch power of the hardware. For example, my copy of MFOG 12 is playing 19x9 at about 1 dan KGS on an old ThinkPad T61. But the same program would be playing at about 3 dan on a workstation with an i7-3770 or i7-4770 in it (I don't think the machine on which Fotland has the bot that is playing on KGS is more powerful that that).

The problem here is that MCTS especially is critical of enough "crunch power" to work at all. Yes levels could be set to "a certain number of playouts" but likely that might be adjusted for "reasonableness check" since who would use something that took half an hour per move? Space and time might also trade off and perhaps some of these small devices lack memory as well as crunch power << except I see no reason why the algorithm would necessarily be greedy for space -- however space limitations could be limiting library sizes >>

Implicitly, iPad, since iOS only runs on iDevices.

I compared the programs on my hardware, for my tastes, not for raw strength, I am not saying one is better than the other.

Also, a T61 of around 2007 is comparable (likely worse) in crunch power than an iPad Air.



Huh!? Pattern libraries, fuseki libraries and transposition tables. With Pachi you can easily kill a device withh less than 1GB RAM, and the "recommended" setting uses 3GB for inner trees.

_________________
Geek of all trades, master of none: the motto for my blog mostlymaths.net

Hi RBerenguel,

Thank you for mentioning this one, I gave it a try and I like this one very much !

[edit]
Did you tryied this one ?
https://itunes.apple.com/fr/app/the-go- ... 24004?mt=8

Beating that one on a 9x9, even in weak mode, is still a very big challenge for me
[/edit]

_________________
Converting the book Shape UP! by Charles Matthews/Seong-June Kim
to the gobook format. last updated april 2015 - Index of shapes, p.211 / 216

I just tried it. It seems decent enough, but it's incredibly slow, compared to CS or Igowin. Also, I'm pretty sure it made a rule mistake with a corner ko in a game I just tried (he won, I needed that ko for a liberty, he played at 1-2 and didn't let me take at 1-1, which was supposedly a legal move.)

_________________
Geek of all trades, master of none: the motto for my blog mostlymaths.net

a -- really? What would a benchmark show for the CPU of an iPad Air? Keep in mind that for practical speed need to consider cache to ram transfers*. How much CPU cache does the CPU of your iPad air have?

b -- Space is of course relative to the history of computers. When I first was involved with computers (60's)even a few K was a big deal. When I again began working with computers in the late 70's a few M was a big deal. What is normal, our ideas of what "lots of ram" means vs "normal amount of RAM" changes over time.

So my "not greedy" meant "not needing more than could be expected as normal". Normal relative to what a "standard computer" could be expected to have. This really hasn't changed all that much since the days when those T61's were new and most new laptops don't have a lot more than the 4 gigs of my old T61's.

Devices like the iPad are powerful communication devices with reasonable computing power for simple tasks. As long as they are being used for their intended purposes they will seem just as strong as an ordinary laptop or desktop. Small, low power consumption devices, not intended for doing a "crunch". Which is fine, most users, even users of laptops and desktops aren't using them to do "crunches" either. But asking a machine to be a decently strong go opponent is asking it to do a "crunch". For any of the available programs to be playing at a few dan at real time speeds (realistic time controls) you'd want at least a machine with as much CPU as say an i7-3770 or i7-4770 in it. But a machine like that isn't out of the question for a home user.



* I learned that the hard way when after retirement I decided to learn C. After getting a decent grasp of the language I decided to tackle a "tuning" exercise but to my surprise even before trying to tune it was running faster than should have been possible (in terms of memory cycles). Didn't take me long to figure out that because I hadn't made the memory usage too large for this to be running entirely within cache.

The A7 CPU has a speed of between 1.4-1.5 Ghz. As much as I can Google the details, so much can anyone else. Here are Anandtech's details. A T61 from around 2007 probably had an Intel Core 2 Duo (according to Lenovo's specs it says Core Duo, but I assume it's a mistake) with between 1.8 and 2.2 Ghz, more likely the former than the latter. According to Lenovo and Intel, it had a 4MB L2 cache and 667-800 MB front bus. An A7 has a very small (64kb instructions and 64kb data, which probably makes branching prediction incredibly fast compared to the Core 2 Duo) L1 cache (which the Core 2 Duo seems to lack,) a 1 MB L2 cache (shared by both computing units) and a 4 MB L3 cache. It may not sound impressive (well, it has L1, which is quite a boost to computing power already) but of course bus technology has changed since 2007: its memory bandwidth is at least 10 GB/s. That's more than 10 times faster than the T61 CPU, and it's not peak speed.

Also, although I guess they are not using it, the A7 has a dedicated GPU. If go software on iOS could take advantage of it for computing something (related to the MCTS algorithms, sure), it could pretty much be faster than non-optimised desktop code.



You keep talking about crunch of this and that CPU, where in no moment I estimated Crazy Stone or Igowin's strengths, AT ALL. So, stop nagging about this. If you are really interested in knowing more details about the speed and level of this programs, either get an iSomething (or an Android device, I think Crazy Stone is also available there) or contact the developers. I have exchanged some emails with the people at Unbalance and they are quite nice.

Also, when I started doing numerical analysis stuff at my degree (2001,) my current iPad Air would have blown well past what I was using. Speed/memory are all matters of the current times. As the adage goes, a current smartphone has much more crunch than the Apollo 11. And that all the Apollo 11 control room, for that matter.

_________________
Geek of all trades, master of none: the motto for my blog mostlymaths.net

haha... a least I won against crazy stones without handicap stones !
ok... I needed 379 score estimation and at least the same amount of undo so maybe that's more crazy stone beating itself... ...
But still... I never did that before what ever I tryed (even with infinite undo )


Usually I still need 5 or 6 stones, but here I was doing some experiment with sansan and I liked the start of the game so I tryed till the end.

_________________
Converting the book Shape UP! by Charles Matthews/Seong-June Kim
to the gobook format. last updated april 2015 - Index of shapes, p.211 / 216