Re: [Computer-go] Mental Imagery in Go - playlist
to err, or not to err... https://www.youtube.com/watch?v=08b0iw3qiAIindex=5list=PL4y5WtsvtduqNW0AKlSsOdea3Hl1X_v-S ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mental Imagery in Go - playlist
introducing... *HALy* https://www.youtube.com/watch?v=UZa2cklrj20index=4list=PL4y5WtsvtduqNW0AKlSsOdea3Hl1X_v-S -- personal website http://sites.google.com/site/djhbrown2/home ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mental Imagery in Go - playlist
Hello, thanks for your feedback. Ingo: Tanja may be the kind of artist who could produce nice drawings of Hajin's mental images, perhaps based on my own crude sketches? It would be unpaid work though... Sorry, but Tanja is a professional. She hs no particular inner relation to the game of Go. So, it would be necessary to pay her. *** Another proposal: Back in 1950 Claude Shannon designed Shannon's Switching Game as a test bed for an analog procedure (instead of digital computing): finding move candidates by electricity flow. One of my students has recently written a program to simulate this. Maybe, I can show some screenshots soon. Richness of a position is shown by many fat edges. Ingo. ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go[http://computer-go.org/mailman/listinfo/computer-go] ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mental Imagery in Go - playlist
However, i have to admit that in 1979 i was a false prophet when i claimed the brute-force approach is a no-hoper for Go, even if computers become a hundred times more powerful than they are now ... I think you are okay: at the point where computers were 100 times quicker than in 1979, monte-carlo was still too slow for anyone to realize its potential. (Fastest supercomputer now is 33.86 petaflop/s, which is approx 10^8 to 10^9 quicker than the fastest in 1979! I think it is about the same ratio for a typical desktop.) :-) Darren ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mental Imagery in Go - playlist
On Tue, Aug 04, 2015 at 10:33:30AM +1000, djhbrown . wrote: However, i have to admit that in 1979 i was a false prophet when i claimed the brute-force approach is a no-hoper for Go, even if computers become a hundred times more powerful than they are now [Brown, D and S. Dowsey, S. The Challenge of Go. *New Scientist* 81, 303-305, 1979.]. I think you are right, though. In my opinion, calling MCTS brute force isn't really fair, the brute force portion really doesn't work and you need to add a lot of smarts both to the simulations and to the way you pick situations to simulate to make things work. -- Petr Baudis If you have good ideas, good data and fast computers, you can do almost anything. -- Geoffrey Hinton ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mental Imagery in Go - playlist
I think you are right, though. In my opinion, calling MCTS brute force isn't really fair, the brute force portion really doesn't work and you need to add a lot of smarts both to the simulations and to the way you pick situations to simulate to make things work. In chess, basic min-max, with an evaluation function that is just the point values for pieces I learnt as a lad (9 for queen, 5 for rook, 3 for knight/bishop, 1 for pawn) would never have beaten Kasparov. (Or could it? I've not followed computer chess closely enough to be sure, but I did hear that Deep Blue was fairly sophisticated software, not just a lot of hardware.) Darren P.S. Isn't brute force the term used to mean that you can see measurable improvements in playing strength just by doubling the CPU speed (and/or memory or other hardware restraint). Alpha-beta with all the trimmings, and MCTS with a good pattern library, both seem to qualify. ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mental Imagery in Go - playlist
On Wed, Aug 5, 2015 at 10:56 AM, Darren Cook dar...@dcook.org wrote: P.S. Isn't brute force the term used to mean that you can see measurable improvements in playing strength just by doubling the CPU speed (and/or memory or other hardware restraint). Alpha-beta with all the trimmings, and MCTS with a good pattern library, both seem to qualify. No, that just means that the solution scales (and brute force solutions tend to scale up quite poorly). https://en.wikipedia.org/wiki/Brute-force_search ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mental Imagery in Go - playlist
on the subject of brutish intelligence, here is a sneak preview of a draft of the script for episode 4 in the series: HALy is an imaginary robot, named after two famous computers: HAL, the antihero of Arthur C. Clarke's wonderful movie 2001: A Space Odyssey, and Haylee, the hero and Secretary General of the International Go Federation. Whereas HAL was made of electronics, Haylee is a real person made of flesh and bone. I'm not being rude to call Haylee a computer, because all human beings - and indeed, all living things, from blue whales to underwater photographers, including you and me and even the bacteria in our guts - are computers. Living computers. Every cell in your body is a miniature computer, made out of what Dennis Bray calls wetware, because the computational machinery of life, made of living plastics like proteins and other stuff, lives in the watery insides of biological cells. HALy's logic is based on what Haylee tells us about when she is playing Go. HAly tries to imagine in it's own head the mental images that Haylee talks about when she is playing. It does this by expressing Haylee's commentaries in the form of symbolic rules; rules that one day a clever computer programmer might be able to turn into computer software so HALy could take the big step from fiction to fact. Let's look at some of Haly's rules: oh, by the way, if HALy were ever able to play a whole game of Go, it would need thousands, possibly millions of rules, but so far i've only thought of a few of them. Here's one, derived from Haylee's explanations in the previous episode of this series: IF i want to play in an empty corner AND opp has some strong stones facing the empty corner on one of the sides next to it THEN look for a joseki that won't give opp a complementary position between its outcome and what he has already on that side This rule is a generalisation of the example Haylee talked about. In that example, she was thinking of where to play in the lower left corner, whilst the lower right corner was occupied by a single opp stone on the hoshi point. HALy can see straightaway whether a corner is empty or not, but how about some of the other qualities in the rule? Like strong, facing and complementary? These need to be worked out; to be thought through. facing is the simplest quality to determine - to keep the explanation simple, let's pretend we are white and opp is black. If the nearest stone along the side is black, it's facing us. Here are some examples of groups that face towards the lower left And some examples of groups that dont. This one doesn't either, but you would hardly call the black stone strong, as it is overshadowed and will have a hard time living. Remember, the rule only applies to strong stones facing the empty corner. So we need to find a way to work out whether stones are strong or not. And right away we are plunged into the forest of complexity, because whether or not a stone is strong depends on whether or not it will live. Tsume-go at the very beginning of fuseki!! How to solve a tricky problem? There are basically two approaches: you can either work hard, or work smart. In Go, working hard means reading it all out - or as much of it as you can. It's the basic strategy used by Monte Carlo search, which operates a bit like a whilrling dervish, flailing around in all directions and relying on a relatively simplstic evaluation function that can at least identify big swings at the end of long sequences, and a prodigious mental energy to read out millions of such sequences. It's a kind of brute force search, which although not exhaustive, is extensive enough to make it hard for its opp to predict what it's going to do next. Monte Carlo players often make bizarre moves that are strikingly dumb but occasionally impressively tricky. The technique has taken the best of them high up in the amateur ranks, far higher than i imagined possible 40 years ago, serving to demonstrate yet again that most of us mere mortals are not as smart as we fondly like to imagine we are! In contrast, working smart means standing on the shoulders of the armies of great players who have gone before you, and by trial and error over the centuries, worked out some general principles that usually work. AI people call such principles heuristics, a Greek word meaning rule of thumb. We use heuristics in our daily lives all the time; and it is possible that using heuristics is the very essence of intelligence. For example, one heuristic used by magicians, footballers, fencers, rugby players, and kangaroos, is the feint. The feint is a brief movement in one direction, immediately followed by a sharp turn and a dodge in the other direction, in order to avoid an onrushing predator. It works because the attacker (or audience member of a magic show) has a brain which, like the brain of the common housefly, is programmed to detect movement and to imagine, as stock market players all too often
Re: [Computer-go] Mental Imagery in Go - playlist
RE: CNNs: They can be, and have been, successfully applied to movies as well. See http://www.cs.cmu.edu/~rahuls/pub/cvpr2014-deepvideo-rahuls.pdf Also, in the first .pdf I linked you, the input layer has a notion of age of the stones. For example, this stone was played 5 moves ago, this one 3 moves ago, etc. So, it is not a strictly static snapshot of a board. In any event, the best performance will probably not come ONLY from CNNs (although its prediction accuracy is surprisingly high), but the marriage of CNNs to monte-carlo tree search, etc. My sense is that we will continue clinging to romantic notions of human intelligence (shapes, proverbs, etc.) until we eventually get ground to dust in a Deep-Blue style competition. Not too long now :) On Sun, Aug 2, 2015 at 9:33 PM, djhbrown . djhbr...@gmail.com wrote: Thanks for the replies to my first message; i looked at the links you supplied and comment on them later in this email. I noticed that Google does not show you the playlist when you look at episode 1 of the series (of currently 3 videos), so you may have missed the second two episodes which are more significant than the first. Here is a link to the playlist: https://www.youtube.com/playlist?list=PL4y5WtsvtduqNW0AKlSsOdea3Hl1X_v-S episode 2 introduces mental images and episode 3 is a conversation between Hajin Lee and me about her thoughts on a couple of moves early in one of her games. It includes my first attempt at picturing her thoughts, both as symbolic information structures and as paint overlays on the game board. My hope is that the former might one day become the basis of symbolic generic heuristic rules that could be used to generate and evaluate move candidates and the latter could evolve into useful instructional materials for people learning the game - so that they can, so to speak, look through the eyes of an expert like Hajin. To these ends, i need the assistance of people with better skills than me at (a) drawing pictures, (b) software and (c) Go. I think that programming is like gymnastics - best done by the young, with their abundance of enthusiasm and energy. I enjoyed programming 50 years ago, but i'm too old in the tooth now to burn midnight oil. Now to your replies: Folkert: Stop is a good start but as you already know, there's a long way to go yet :) Steven: I expect there is a future for CNN's in recognising static images, but my gut feel is that a position in a Go game is more like one frame of a movie; as such, it requires a technology that can interpret dynamic images - maybe work being done in automatous car driving can contribute something useful to Go playing? Nevertheless, I was surprised by the many humanlike moves of DCNNigo on KGS (until it revealed its brittleness). To be sure, drawing upon the moves of experts is one way of gaining expertise, but my feeling is that one should try to abstract the position - to generalise from the examples - so that general knowledge can be formed and applied to novel situations. It may be that a CNN arguably does do some kind of generalisation - but can it, for example, characterise something as basic as the waist of a keima? Ingo: Tanja may be the kind of artist who could produce nice drawings of Hajin's mental images, perhaps based on my own crude sketches? It would be unpaid work though... I liked Fuego's and Jonathan's territory pictures, which reminded me of Zobrist's early work on computing influence. [Albert Zobrist (*1969*). *A Model of Visual Organisation for the Game of Go*. Proceedings of the Spring Joint Computer Conference, Vol. 34, pp. 103-112.] However, whereas being able to picture influence and territory is one of my objectives, i want to try to picture the richness of what Hajin (aka Haylee) sees rather than the result of a primitive computation. For example, at 10:24 in episode 3, she points out that when black is on J4 instead of K4, there is an opening in black's lower side for white to invade. This tiny gap makes all the difference to the dynamic meaning of the position a few moves prior (ie whether it is sensible for white to approach Q3 at Q5). One of the major influences on my own thinking about Go programming is the seminal work Thought and Choice in Chess by Adriaan de Groot which i reckon is well worth a read by anyone interested in programming Go https://books.google.com.au/books?id=b2G1CRfNqFYCpg=PA99 --- personal website http://sites.google.com/site/djhbrown2/home ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mental Imagery in Go - playlist
Thanks for the link to the CMU CNN paper, Steven, which was very interesting. I noted with some pleasure that they included a fovea stream - although maybe that is a bit of a misnomer, as whereas animal foveas roam around the image, building (i think) a symbolic structural description of the picture, theirs was fixed in the middle. I wonder whether a roaming fovea CNN could be a successful group connectedness classifier? I can envisage the fovea being moved around by a higher-level routine that uses a symbolic description of the game situation to identify which areas/groups it wants it to investigate. Incidentally, i'm unconvinced that including an age of stone feature is valuable, because although the future is dynamic, the past is set in stone (sic); Go teachers sometimes talk about tewari analysis to demonstrate when an old stone becomes inefficiently placed by a certain line of play. As to romantic notions of human superiority, i personally feel that such opinions are not so much romantic as hubristic - or perhaps paranoid! However, i have to admit that in 1979 i was a false prophet when i claimed the brute-force approach is a no-hoper for Go, even if computers become a hundred times more powerful than they are now [Brown, D and S. Dowsey, S. The Challenge of Go. *New Scientist* 81, 303-305, 1979.]. Back in those days, i never imagined that something so blind as Monte-Carlo would become more perceptive than even my weak eye, let alone being able to defeat a pro (albeit with a 5-stone handicap), as Zen just did on KGS. By the way, i've long since lost my paper copy of my paper; you have access to an academic library - would you be able to retrieve and scan a copy of it, just for my nostalgia? -- personal website http://sites.google.com/site/djhbrown2/home ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] Mental Imagery in Go - playlist
Thanks for the replies to my first message; i looked at the links you supplied and comment on them later in this email. I noticed that Google does not show you the playlist when you look at episode 1 of the series (of currently 3 videos), so you may have missed the second two episodes which are more significant than the first. Here is a link to the playlist: https://www.youtube.com/playlist?list=PL4y5WtsvtduqNW0AKlSsOdea3Hl1X_v-S episode 2 introduces mental images and episode 3 is a conversation between Hajin Lee and me about her thoughts on a couple of moves early in one of her games. It includes my first attempt at picturing her thoughts, both as symbolic information structures and as paint overlays on the game board. My hope is that the former might one day become the basis of symbolic generic heuristic rules that could be used to generate and evaluate move candidates and the latter could evolve into useful instructional materials for people learning the game - so that they can, so to speak, look through the eyes of an expert like Hajin. To these ends, i need the assistance of people with better skills than me at (a) drawing pictures, (b) software and (c) Go. I think that programming is like gymnastics - best done by the young, with their abundance of enthusiasm and energy. I enjoyed programming 50 years ago, but i'm too old in the tooth now to burn midnight oil. Now to your replies: Folkert: Stop is a good start but as you already know, there's a long way to go yet :) Steven: I expect there is a future for CNN's in recognising static images, but my gut feel is that a position in a Go game is more like one frame of a movie; as such, it requires a technology that can interpret dynamic images - maybe work being done in automatous car driving can contribute something useful to Go playing? Nevertheless, I was surprised by the many humanlike moves of DCNNigo on KGS (until it revealed its brittleness). To be sure, drawing upon the moves of experts is one way of gaining expertise, but my feeling is that one should try to abstract the position - to generalise from the examples - so that general knowledge can be formed and applied to novel situations. It may be that a CNN arguably does do some kind of generalisation - but can it, for example, characterise something as basic as the waist of a keima? Ingo: Tanja may be the kind of artist who could produce nice drawings of Hajin's mental images, perhaps based on my own crude sketches? It would be unpaid work though... I liked Fuego's and Jonathan's territory pictures, which reminded me of Zobrist's early work on computing influence. [Albert Zobrist (*1969*). *A Model of Visual Organisation for the Game of Go*. Proceedings of the Spring Joint Computer Conference, Vol. 34, pp. 103-112.] However, whereas being able to picture influence and territory is one of my objectives, i want to try to picture the richness of what Hajin (aka Haylee) sees rather than the result of a primitive computation. For example, at 10:24 in episode 3, she points out that when black is on J4 instead of K4, there is an opening in black's lower side for white to invade. This tiny gap makes all the difference to the dynamic meaning of the position a few moves prior (ie whether it is sensible for white to approach Q3 at Q5). One of the major influences on my own thinking about Go programming is the seminal work Thought and Choice in Chess by Adriaan de Groot which i reckon is well worth a read by anyone interested in programming Go https://books.google.com.au/books?id=b2G1CRfNqFYCpg=PA99 --- personal website http://sites.google.com/site/djhbrown2/home ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
