Re: [Computer-go] GCP passing on the staff ...
On 29.01.2019 18:53, uurtamo wrote: it's [...] about an insane need to keep sente. my only takeaway other than reading out fights way way way in advance. I can confirm the necessity for keeping sente with respect for the endgame but would not be surprised it to also apply during opening and middle game. One of the greatest weaknesses of my pupils in the kyus is not to play all their sentes (other than privileges preserved for ko threats or liberties) before gotes. From my study, research of and book writing on the endgame during the previous 2.5 years, I have realised the importance of distinguishing gote from sente even if their difference is only a fraction of a point (but it can be up to ca. 5 points per local decision) and of exceptionally playing gote instead of sente or vice versa depending on the global context. Every small mistake in evaluation about playing too long locally etc. amounts to a large total amount when all mistakes accumulate. Programs would notice such implicitly due to their smaller winning chances when making too many such mistakes. Reading out fights in advance very deeply I have only noticed a few times during programs' play but, of course, you are right. I simply have not studied their deep reading carefully enough to witness more incidents. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 30.10.2017 19:22, Pierce T. Wetter III wrote: this car and this child In Germany, an ethics commission has written ethical guidelines for self-driving cars with also the rule to always prefer avoiding casualties of human beings. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 28.10.2017 11:13, Petri Pitkanen wrote: Exactly verbalized rules lose to pure analysis power. (I think with "verbalised" you mean "codified in writing", with "pure analysis power" you mean "volume of reading, calculation, sampling or NN processing".) Rules are not meant to win or lose against "pure analysis power" but to use it when necessary and unavoidable, e.g., tactical reading when clarifying L+D status. A rule can be "Consider an attack if the L+D status is 'unsettled'" but also tactical reading determines that status. Human intuition is trained with endless repetition. IMO, intuition does not exist; it is nothing but an excuse for not understanding subconscious or currently unobservable thinking yet. Can we speak of human subconscious thinking, please? Subconscious thinking can be trained by learning rules, practising problems etc. Conscious, explicit thinking can be trained by learning rules, practising problems etc. So what do you want to say? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 27.10.2017 13:58, Petri Pitkanen wrote: doubt that your theory is any better than some competing ones. For some specialised topics, it is evident that my theory is better or belongs to the few applicable theories (often by other amateur-player researchers) worth considering. For a broad sense of "covering every aspect of go theory", I ask: what competing theories? E.g., take verbal theory teaching by professional players and they say, e.g., "Follow the natural flow of the game". I have heard this for decades but still do not have the slightest idea what it might mean. It assumes meaning only if I replace it by my theory. Or they say: "Respect the beauty of shapes!" I have no idea what this means. A few particular professional players have reasonable theories on specific topics and resembling methodical approach occurring in my theories. So what competing theories do you mean? The heritage of professional shape examples? If you want to call that theory. As I do know people who are stronger than you and are using different framework. Yes, but where do they describe it? Almost all professional players I have asked to explain their decision-making have said that they could not because it would be intuition. A framework that is NOT theory. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 26.10.2017 08:52, Petri Pitkanen wrote: Unfortunately there is no proof that you principles work better than those form eighties. No computer-go proof. There is evidence in the form of my playing strength: with the principles "from the eighties", I got to circa 1 kyu. L+D reading practice etc. made me 3 dan. Afterwards, almost the only thing that made me stronger to 5 dan and then further improved my understanding was the invention of my own principles. My principles etc. also work for (an unknown fraction of) readers of my books and for a high percentage of my pupils but I cannot compare what the effect on them would have been if instead they would only have learnt the principles "from the eighties". I do, however, know that my principles provide me with very much more efficient means of teaching contents compared to using the principles "from the eighties". The principles "from the eighties" and my principles can be compared with each other. IMO, such a comparison is shocking: the principles "from the eighties" are very much weaker on average and altogether convey very much less contents. Nor there is any agreement that your pronciples form any improvement over the old ones. Only time constraints prevent me from doing an extensive comparison and so better support formation of an agreement. What is missing that I doubt that you can verbalise your go understanding to degree that by applying those principles I could become substantially better player. Different players are different. So different that some players claim to only learn from examples. Therefore, I cannot know whether you are a player who could learn well from principles etc. - My reading skills would not get any better Do you say so after having learnt and invested effort in applying the contents of Tactical Reading? Regardless of the possible impact of that book, a great part of reading skill must be obtained by reading practice in games and problem solving. If your reading is much weaker than your knowledge of go theory, then it may be the case that almost only reading practise (plus possibly reading theory about improving one's reading practice) can significantly improve your strength at the moment. - your principles are more complex than you understand. I do not think so:) Much of you know is automated to degree that it is subconsciousness information. From ca. 10 kyu to now, especially from 3 dan to now, I have reduced the impact of my subconscious thinking on my go decision-making and replaced it by knowledge, reading and positional judgement based on knowledge and reading. The still remaining subconscious thinking is small. Most of my remaining mistakes are related to psychology or subconscious thinking, when necessary because of explicit knowledge gaps or thinking time constraints. Transferring that information if hard. Transferring it from principles etc. to code - yes. If you can build Go bot about KGS 3/4dan strength Using my approach, I expect several manyears, which I do not have for that purpose. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] I present my apologizes to Robert Jasiek
Accepted, thank you! -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 26.10.2017 13:52, Brian Sheppard via Computer-go wrote: MCTS is the glue that binds incompatible rules. This is, however, not what I mean. Conflicting principles (call them rules if you like) must be dissolved by higher order principles. Only when all conflicts are dissolved should MCTS be applied. What you describe has been used with success and better success than I expect what my knowledge-pure approach can currently achieve. But MCTS as glue for conflicting principles has also run into a boundary. I want to see that boundary surpassed by my pure approach. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 24.10.2017 20:19, Xavier Combelle wrote: totally unrelated No, because a) software must also be evaluated and can by go theory and b) software can be built on exact go theory. That currently (b) is unpopular does not mean unrelated. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 24.10.2017 16:45, Xavier Combelle wrote: I don't understand what you mean by go-theorical aspects. Go theory is an ambiguous term and means everything from informal ("Starting with a standard corner move can't be wrong.") via principle ("Usually, defend a weak important group.") to formal ( https://senseis.xmp.net/?CycleLaw ). and especially when applying to computer-go. Relating computer play / algorithms to go theory or vice versa adds another layer of difficulty indeed. To my knowledge the only theoretical (in a mathematic meaning of theoretical) approach of go is combinatorial theory and it leads to very few knowledge. Other mathematical theory with practical relevance is related to capturing races (see Capturing Races 1 - Two Basic Groups, Thomas Wolf's papers etc., endgame (e.g., http://home.snafu.de/jasiek/kodame.pdf and google for related proofs) or will be published by me later (will be quite a lot and have practical relevance, but you need to be patient). Research in mathematical go theory requires much time because exactness is often necessary and proving can be tricky. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 24.10.2017 11:45, David Ongaro wrote: very seldom saw a discussion with Robert lead to anything. (You seem to only refer to discussion on this mailing list.) Apart from this being a discussion about one particular person, let me ignore this for a moment: In the current time, computer-go discussion and research has a very high percentage of people discussing the side of mainly programs and programming but I belong to the very low percentage of people discussing mainly go-theoretical aspects of computer-go. With a higher percentage of the latter, there would also be more discussions resulting to something. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 23.10.2017 19:15, Xavier Combelle wrote: [personal attack deleted] Did you already encounter a real game with "disturbing life kos or anti-sekis" and especially "ladders (...) beyond 250 moves" ? If not how do you believe that Alphago would learn how to manage such situations. Dave Dyer wrote: I wonder how alphago-0 treats the menagerie of special positions, such as bent 4 in the corner, thousand year ko, rotating ko, etc. uurtamo wrote: > It will be interesting to realize that those specialized positions > (thousand-year-ko, bent 4) are actually a microscopic issue in > game-winning. The exceptional cases may be rarities in practical play but not all are that rare. E.g., I have had two games in roughly 40,000 ending with a double ko seki. Already one "rarity" occurring occasionally means that all rarities occur more often in practice. Therefore they do have practical relevance. Quite like a white truck is relevant and not to be confused with the sky, or an AI car can kill (which has happened because of such a "rarity"). In go, the consequences or misjudging "rarities" are just up a lost game, but this is the very purpose - avoiding lost games by avoiding errors. Rarities are good test samples for checking whether an AI program avoids errors in non-standard situations. The same must be studied for standard situations, whose deeper details can also lead to errors. Not because a standard by itself would be difficult but because the deeper details increase complexity and this can lead to errors. Studying the standards and identifying errors in their deeper details can be difficult. E.g., we see AlphaGo (Zero) invading and living in a large moyo or not invading and wonder why. Part of the answer would be: invading and living is impossible. Studying this is complex because it involves deep reading for the standard case of a moyo and the question of invading it. The "rarities" are infrequent but can be good test tools because distinguishing correct from wrong play can be easy if a rarity's behaviour is understood well. The standards are frequent but often not the best test tools because many standards interact with each other and they all depend on deep reading and exact positional judgement. I cannot know if AlphaGo Zero has already learnt how to play in (some) rarities (those that can be solved earlier than the constant game end rule; e.g., we cannot test 4 octuple kos), will learn it or would not be able to learn it - but I want to know. In particular, because I want to know which errors AlphaGo Zero does make. I want to know this for go and for the general AI project. Avoiding errors is essential for both. I do not fall into the illusion that AlphaGo Zero would be the perfect player but expect that it can make errors at any (unexpected) time. We need to understand what causes errors, how frequent they are and what most extreme consequences they can have. Rarities are one very good study tool for this purpose. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
On 23.10.2017 14:05, Jim O'Flaherty wrote: Couldn't they be useful as part of a set of training data for newly trained engines and networks? All the millions of games would be very useful for many purposes. E.g., I want to know whether the reconstructed knowledge includes such basic things as terminal positions with disturbing life kos or anti-sekis, whether ladders are recognised beyond 250 moves etc. Not to mention non-go applications. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] AlphaGo Zero SGF - Free Use or Copyright?
AlphaGo Zero games are available as zipped SGF from Deepmind at http://www.alphago-games.com/ For earlier AlphaGo games, I have seen statements from Deepmind encouraging free use (presuming stating origin, of course) so that the games may be commented etc. I cannot find a similar statement from Deepmind for the published AlphaGo Zero games. Are they for free use or copyrighted? I hope the former so everybody including Deepmind can see more commentaries. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
On 20.10.2017 21:12, uurtamo . wrote: do something like really careful experimental design across many dimensions simultaneously (node weights) and several million experiments -- each of which will require hundreds if not tens of thousands of games to find the result of the change. Worse, there are probably tens of millions of neural nets of this size that will perform equally well (isomorphisms plus minor weight changes). So many changes will result in no change or a completely useless game model. It is possible that things turn out as complex as you describe... "modeling through human knowledge" neural nets doesn't sound like a sensible goal ...but I am not convinced. Researchers in the human brain's thinking keep their optimism, too. Nevertheless, alternative approaches can be imagined. E.g., while building a neural net of eventually great strength also build in its own semantic interpretator, semantic verificator (including exclusion of errors as far as computationally possible) and translator between internal structure and human (or programming) language representation. I do not know if such dynamic self-representations of neural nets have already been described but if not this would be an interesting research topic. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Subject: Re: AlphaGo Zero
On 20.10.2017 16:44, Hendrik Baier wrote: Where is the respect and the appreciation for other people's groundbreaking work without immediately having to make the discussion about your own research, or otherwise derailing it into the irrelevant or fantastical? Instead of joining your meta-discussion, let me point out some motivation: - Research should also proceed after groundbreaking work. - Research should not be isolated but different research approaches can be combined. Research is not a one-way street. - Research in go also serves as a model for research in other or more general fields, such as generalised AI, which includes use of AI to provide a) interchange with semantics of human domains or b) assistance for or replacement of human activities, such as car driving, which profit from avoiding errors. Therefore I discuss these aspects. - Ethical questions are becoming increasingly important in view of the fast progress of AI. Surely you are aware that Deepmind knows this. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
On 20.10.2017 15:07, adrian.b.rob...@gmail.com wrote: 1) Where is the semantic translation of the neural net to human theory knowledge? As far as (1), if we could do it, it would mean we could relate the structures embedded in the net's weight patterns to some other domain -- The other domain can be "human go theory". It has various forms, from informal via textbook to mathematically proven. Sure, it is also incomplete but it can cope with additions. The neural net's weights and whatnot are given. This raw data can be deciphered in principle. By humans, algorithms or a combination. You do not know where to start? Why, that is easy: test! Modify ONE weight and study its effect on ONE aspect of human go theory, such as the occurrance (frequency) of independent life. No effect? Increase the modification, test a different weight, test a subset of adjacent weights etc. It has been possible to study semantics of parts of DNA, e.g., from differences related to illnesses. Modifications on the weights is like creating causes for illnesses (or improved health). There is no "we cannot do it", but maybe there is too much required effort for it to be financially worthwhile for the "too specialised" case of Go? As I say, a mathematical proof of a complete solution of Go will occur before AI playing perfectly;) So far neural nets have been trained and applied within single domains, and any "generalization" means within that domain. Yes. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
On 20.10.2017 09:38, Xavier Combelle wrote: What is currently named nanorobot is simply hand assembled molecules which have mechanical properties and need huge framework to be able simply move. Sure. But we must not wait until such a thing exists. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
On 20.10.2017 07:10, Petri Pitkanen wrote: >> 3) Where is the world-wide discussion preventing a combination of AI >> and (nano-)robots, which self-replicate or permanently ensure energy >> access, from causing extinction of mankind? 3) Would it be a bad thing? All thing considered, not just human point of view Have you realised the potential of one successful self-duplication of a nano-robot? Iterate and the self-replicating nano-robots might dissolve the planet earth into elementary particles. Now discuss whether that might be good or bad. Not good for animals or plants, to start with. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo Zero
So there is a superstrong neural net. 1) Where is the semantic translation of the neural net to human theory knowledge? 2) Where is the analysis of the neural net's errors in decision-making? 3) Where is the world-wide discussion preventing a combination of AI and (nano-)robots, which self-replicate or permanently ensure energy access, from causing extinction of mankind? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] agz -- meditations
On 19.10.2017 20:13, Richard Lorentz wrote: Silver said "algorithms matter much more than ... computing". Hassabis estimated they used US$25 million of hardware. Today, it seems 4 TPU cost US$25 million. In 5 or 10 years, every computer might have its 4-TPU-chip costing $250, if not $25. At least, I hope. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] AlphaGo's Endgame Mistakes
Reading Invisible, it is apparent that AlphaGo makes score-related mistakes in the endgame, ko fights or virtual ko fights (read: wasting ko threats) occurring during the early endgame if AlphaGo wins nevertheless. So we cannot say yet that they would be win-related (or winning-probability-related) mistakes. AlphaGo plays better endgame if it needs to. The score-related mistakes are easily explained in terms of traditional human go theory or more clearly in terms of formal go theory using the score-related view (larger score is better than smaller score in perfect play with perfect information). So far, it seems unknown whether AlphaGo might also make some of those mistakes when its win is still unclear (winning probability near 50%). Improving AlphaGo's play WRT to the score-related mistakes seems straightforward: first create moves as currently, then dynamically iterate komi increments for specific positions during the games and create a second instance of AlphaGo modified due to its improved play with tougher komi. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Deep Blue the end, AlphaGo the beginning?
On 17.08.2017 21:35, Darren Cook wrote: The machine-learning methods AlphaGo uses are applicable to practically anything." They (alone) are not good for guaranteed avoiding of mistakes (as is achieved by theorems), or for guaranteed permanent execution by their software, hardware, power supply and circumstances of natural forces. They (alone) are not enough for GAI also because many applications need some interface between the machine-learning methods and the domain-specific requirements. They are inapplicable to what is proven to be, or their application can be unpredictable in some sense for what is proven to be non-deterministic. They can pretend to simulate but cannot replace religion, meta-physics, ethics, emotions. However, it may be an open debate whether they can assume self-consciousness some time. If we give them too much power (incl. self-regulated access to physical ressources), they can render extinct mankind. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Alphago and solving Go
On 09.08.2017 20:50, John Tromp wrote: The number of games is at most 361^#positions. This misses passes, rules distinguishing situations etc. and infinite sequences under some rulesets. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Notes from the Asilomar Conference on Beneficial AI
On 15.02.2017 09:42, adrian.b.rob...@gmail.com wrote: if humans tried to play that way (or that way at those times), it wouldn't turn out very well because working memory or other requirements to pull it off are just too high. My reply to this is the same as before. (Some humans can play like this successfully.) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Notes from the Asilomar Conference on Beneficial AI
On 10.02.2017 12:56, adrian.b.rob...@gmail.com wrote: AlphaGo is playing moves and styles that all human masters had dismissed as stupid centuries ago." we may learn little more than what mathematicians learn when a computer-assisted proof consisting of several hundred pages is generated for a conjecture like Fermat's last theorem. The "novelties" AlphaGo shows are not new to humans. Just new to the vast majority of human players having followed mainstream strategy and style. Players with highly creative, flexible styles have not been surprised by AlphaGo's moves - only by AlphaGo's "reading" depth. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo rollout nakade patterns?
On 31.01.2017 16:39, Gian-Carlo Pascutto wrote: http://senseis.xmp.net/?BasicLivingEyeShapes Warning: these do not include any living eye shapes with inside stones, nor specialities on the edge or in the corner. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo rollout nakade patterns?
On 23.01.2017 19:00, terry mcintyre via Computer-go wrote: nakade involves creating a shape (such as three in a row or a bulky five) such that, if captured, it would only form one eye, given the proper placement. Nakade has been defined (e.g., several times by me) reasonably well, but for computer purposes some sort of simplifying (implicit) definition is often necessary according to a study purpose. (Your attempt is too naive, worse than ca. 68 years ago. See my texts for progress or else be naive with determination, i.e., keep things simple strictly without any "such that", "would", "eye", "proper", "involves".) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] ADMIN: Lists Have Moved
I suppose we still send messages to computer-go@computer-go.org Yes, if I receive this message from the list. I suppose nothing has changed about the procedures and email addresses for (un)subscription. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Are the AlphaGols coming?
On 09.01.2017 23:03, David Ongaro wrote: > decisions are normally made subconsciously seconds before we get > aware of them Essentially nothing is known how to interpret such neurological findings. It is (usually) not like the universe was forcing me unexpected subconscious thinking into my conscious mind. My topic-dependent thinking occurs because I want to be busy thinking about the topic for a long time (such as successive minutes or hours - not seconds as in the tests - during a go game). In such a thinking context, both subconscious and conscious thinking related to the topic occur with countless interactions in both directions (and even occasional level changes of subconscious pieces accessible as conscious, but this is not so interesting, it is like reading in assembler;) ) Now, if some test claims to observe that subconscious thinking preceded conscious thinking, this is like making assumptions of excluding parts of conscious thinking. As if you wanted to deceive Heisenberg's uncertainty relation. Maybe it does play a relevant role in brains. Observation affects perception. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Are the AlphaGols coming?
On 09.01.2017 07:19, David Ongaro wrote: >> accurate positional judgement you also rely on “feelings” otherwise you wouldn’t be able to survive. In my go decision-making, feelings / subconscious thinking (other than usage of prior sample knowledge, such as status knowledge for particular shapes) have an only marginal impact. For me, they serve as a preselection filter besides my used methodical preselection filters. In blitz, the impact is larger when time is insufficient for always using the methodical ones. Another factor is my pruning of reading. I would not describe it as "feelings / subconscious thinking" but as "prune according to knowledge / principles AFA time allows, otherwise call my mental random generator for deciding what else to prune". I.e., it is a conscious calling of random for particular purposes. Instead of suspecting feelings, read my books - Positional Judgement 1 - Territory - Positional Judgement 2 - Dynamics to better understand why my accurate positional judgement does not need feelings / subconscious thinking. Even in ca. 1/3 of my blitz (10' SD) games, I can apply it (less frequently per game, OC). About the only relevant feeling permitted in my go is a contribution to the decision on my first move as Black, which may also depend on my mood (besides opponent, komi, time, knowledge). 18+ years ago, I used feelings and the like for quite a few decisions during the middle game and (early) endgame. Decision by feelings led to low winning probability so I decided to overcome them by creating much more profound theory, which improved my play and enabled(!) me to survive (to use your words) as a go teacher and go book author. > Mathematically (the approach you seem yourself constrain into) Reasoned decision-making need not always be low-level / mathematical. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Our Silicon Overlord
On 07.01.2017 16:33, Jim O'Flaherty wrote: I hope you get access to AlphaGo ASAP. More realistically, I (we) would need to translate the maths into algorithmic strategy then executed by a program module representing the human opponent. Such is necessary because no human can remember everything to create a legal superko sequence of over 13,500,000 moves or have the mere stamina to perform it. (Already just counting to 1 million is said to take 3 weeks without sleep...) Anyway,... > exploring AI weaknesses ...this is a major objective. E.g., we do not want AI driven cars working right most of the time but sometimes killing people because the AI faces situations (such as a local sand storm or a painting on the street with a fake landscape or fake human being) outside its current training and reading. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Our Silicon Overlord
On 06.01.2017 23:37, Jim O'Flaherty wrote: into a position with superko [...] how do you even get AlphaGo into a the arcane state in the first place, I can't in practice. I have not provided a way to beat AlphaGo from the game start at the empty board. All I have shown is that there are positions beyond AlphaGo's capabilities to refute your claim that AlphaGo would handle all positions well. Hui and Lee constructed positions with such aspects: Hui with long-term aji, Lee with complex reduction aji. Some versions of AlphaGo mishandled the situations locally or locally + globally. The professional players will be open to all sorts of creative ideas on how to find weaknesses with AlphaGo. Or the amateur players or theoreticians. Perhaps you can persuade one of the 9p-s to explore your idea of pushing the AlphaGo AI in this direction. Rather I'd need playing time against AlphaGo. IOW, we are now well outside of provable spaces For certain given positions, proofs of difficulty exist. Since Go is a complete-information game, there can never be a proof that AlphaGo could never do it. There can only ever be proofs of difficulty. mathematical proof around a full game From the empty board? Of course not (today). We cannot formally prove much simpler models, Formal proofs for certain types of positions (such as with round_up(n/2) n-tuple kos) exist. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Our Silicon Overlord
On 05.01.2017 17:32, Jim O'Flaherty wrote: I don't follow. 1) "For each arcane position reached, there would now be ample data for AlphaGo to train on that particular pathway." is false. See below. 2) "two strategies. The first would be to avoid the state in the first place." Does AlphaGo have any strategy ever? If it does, does it have strategies of avoiding certain types of positions? 3) "the second would be to optimize play in that particular state." If you mean optimise play = maximise winning probability. But... optimising this is hard when (under positional superko) optimal play can be ca. 13,500,000 moves long and the tree to that is huge. Even TPU sampling can be lost then. Afterwards, there is still only one position from which to train. For NN learning, one position is not enough and cannot replace analysis by mathematical proofs ALA the NN does not emulate mathematical proving. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Are the AlphaGols coming?
On 06.01.2017 03:36, David Ongaro wrote: Two amateur players where analyzing a Game and a professional player happened to come by. So they asked him how he would assess the position. After a quick look he said “White is > leading by two points”. The two players where wondering: “You can count that quickly?” Usually, accurate positional judgement (not only territory but all aspects) takes between a few seconds and 3 minutes, depending on the position and provided one is familiar with the theory. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Our Silicon Overlord
On 04.01.2017 22:08, "Ingo Althöfer" wrote: humanity's last hope The "last hope" are theoreticians creating arcane positions far outside the NN of AlphaGo so that its deep reading would be insufficient compensation! Another chance is long-term, subtle creation and use of aji. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Very interesting presentation of Fan Hui about AlphaGo
Fan Hui's presentation was very disappointing, except for the statement that there will be a website with commentaries on AlphaGo games. The other possibly interesting information that its play suggests one might play everything during the opening was not new to me, who I have applied this insight since 2004 in my games. I understand how hard it was to get a visa and that Aja Huang failed to get one. The Q section was missing so in particular none of my questions below could be answered yet. I do not give up hope, so here my questions are again: 1) When will there be a commercial program for an ordinary PC having AlphaGo's strength? 2) During the Lee Sedol match, how would disconnection from the internet, hardware failure, operating system failure or software failure have been handled? Would AlphaGo have lost such a game? 3) How well does AlphaGo play in exceptional positions designed by theoreticians, such as complex capturing races or multiple kos with long cycles? 4) Is it possible to relate AlphaGos data structures with human go theory? Can we learn go theory from them (e.g., how to play ko threats)? Can AlphaGo learn from human go theory (e.g., how to play ko threats)? 5) Why exactly did AlphaGo improve so fast? 6) When will a program only self-learning go from scratch by itself reach AlphaGo's strength? 7) What does AlphaGo mean for the progress of General Artificial Intelligence? 8) Suppose one day there is General Artificial Intelligence and it is used for a task involving responsibility, such as health care or car driving. Who has which legal responsibility if the General Artificial Intelligence makes mistakes killing people? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] (no subject)
On 01.04.2016 02:22, djhbrown . wrote: kogo is great for corner openings Kogo contains many mistakes. Too many kyus got their hands on it. It would be better to spend 3+ weeks using kombilo on GoGoD and create a new joseki tree. A summary of such an effort (with some interesting, additional, old variations) you find in Joseki 3 - Dictionary. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] (no subject)
On 31.03.2016 16:54, Bill Whig wrote: Wouldn't you agree that a lot of people (most?) might might advance more swiftly > with move suggestions rather than text that they have to work through like a textbook? I say the opposite. Move suggestions without any additional information are meaningless. There are different learning styles, among them learning by example and learning by theory. Learning by example is not "learning by move suggestions" but is "learning by example moves, example sequences or example positions etc. TOGETHER WITH SOME additional information, such as move together with shape, move together with the positional context, move together with tactical reading and decision-making etc. I have never seen or heard of anybody learning ONLY by example or ONLY by theory. I see all strong players, including those preferring learning by example, having a good explicit or implicit textbook knowledge. "swiftly" versus "have to work through" creates a false impression. Learning by example requires very many examples. Learning by textbook theory requires learning an only intermediate number of theoretical bits. (Learning by move suggestions, i.e., without any additional information, is the slowest. It becomes efficient only for AI having the compuational power to simulate thousands of man-years of learning.) > I haven't ready any of your books, but I've read a > few Go books and most of them do not do justice to the complexity of the game. So you have read the wrong books. "Move suggestions" would invite the student to think in more creative, and effective, ways > than I have seen put forth in any book thus far. 1) Even ABC move suggestion books do not just suggest moves but also provide positional context. 2) You have a totally wrong view on creative, effective thinking if you do not consider the possibility of having it when studying theory. 3) You have read the wrong books. > Most say that the first thing that one should do after learning joseki is to forget it... . It is possible that this nonsense is still believed by a majority. Overcome it, see e.g. a report suggesting differently: http://www.lifein19x19.com/forum/viewtopic.php?f=57=12951 A better advice is: improve while studying joseki by understanding them and their theory. The bad proverbs are about learning josekis without understanding: learn, forget, repeat; this effort can demote a player's strength indeed;) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] new challenge for Go programmers
On 31.03.2016 03:52, Bill Whig wrote: If the program would merely output 3-5 suggested positions, that would probably suffice. Even an advanced beginner, such as myself, could I believe, understand why they are good choices. Just having the "short list" would probably be quite an educational tool! It would probably even help teach joseki. No. Joseki learning requires much more than move suggestions. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] new challenge for Go programmers
djhbrown, Even from a pure playing stronger perspective, it is not game over yet because there is no guarantee yet for always avoiding sudden entering of holes of bad play, verification by reading is missing and there is no optimisation for better score when winning the game anyway. For other AI applications, this means that there is no guarantee of unexpected bad actions, such as accidentally killing people. As I have explained, shortly after the sacrifice squeeze in game 5 AlphaGo had a winning position. Therefore, currently one should not call the squeeze a mistake. A more cateful study of the few moves after the squeeze is necessary. You mention several outdated principles and concepts, whose insufficiency I have explained elsewhere. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] new challenge for Go programmers
On 30.03.2016 16:58, Jim O'Flaherty wrote: My own study says that we cannot top down include "English explanations" of how the ANNs (Artificial Neural Networks, of which DCNN is just one type) arrive a conclusions. "cannot" is a strong word. I would use it only if it were proven mathematically. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Nice graph
On 26.03.2016 06:15, Robert Jasiek wrote: 9d does not exist. I mean as a real world rank. Of course, there are servers in which ranks are derived from ratings. E.g., KGS 9d can mean everything from real world 3d [sic!] to 9p. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Nice graph
On 26.03.2016 01:23, Rémi Coulom wrote: http://i.imgur.com/ylQTErVl.jpg 9d does not exist. 8d is rare and may as well be translated to the very strongest 7d. EGF 7d means up to ca. 5p. Korean 7d might be stronger. EGF 6d means up to ca. 1p. Korean 6d might be stronger. Korean 5d means ca. EGF 5d - 6d. 9p has a great range in itself. Rating systems can have a problem of running away ratings at the top. Self-played ratings might not be significant. 5 games are not enough to assign a secure rank to AlphaGo v. 18. So, no, the graph is not nice. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Congratulations to AlphaGo (Statistical significance of results)
On 23.03.2016 15:32, Petr Baudis wrote: these are beautiful posts. https://massgoblog.wordpress.com/2016/03/11/lee-sedols-strategy-and-alphagos-weakness/ Before you become too excited, also read my comments on the commentary: http://www.lifein19x19.com/forum/viewtopic.php?p=200539#p200539 -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] AlphaGo - Lee Sedol Game 5,Move 68
For those of you not reading forums but interested in a positional judgement after the tower squeeze sacrifice in game 5, an SGF is attached inline. Conclusion: afterwards White has at least a small lead, so AlphaGo's strategy was superb. -- robert jasiek (;SZ[19]GM[1]CA[UTF-8]FF[4]ST[2]AP[GOWrite:2.3.48]PM[2]RU[Chinese]FG[259:]C[Commentary by Robert Jasiek Terms and principles are explained in the book Positional Judgement 1 - Territory. The file contains markup, same-move variations, same-pass variations, nodes with SGF diagrams and markup only, variations with SGF diagrams only. You need a gppd SGF editor to see all diagrams, variations and markups.]BR[9p]PW[AlphaGo]RO[5]DT[2016-03-15]GN[ ]KM[7.50]RE[W+R]PB[Lee Sedol] ;B[qd] ;W[dd] ;B[pq] ;W[dp] ;B[oc] ;W[po] ;B[qo] ;W[qn] ;B[qp] ;W[pm] ;B[nq] ;W[qe] ;B[pe] ;W[qf] ;B[rd] ;W[pf] ;B[ql] ;W[oe] ;B[pl] ;W[ol] ;B[om] ;W[ok] ;B[nm] ;W[qj] ;B[rn] ;W[gq] ;B[cf] ;W[fc] ;B[bd] ;W[ch] ;B[dh] ;W[di] ;B[dg] ;W[cc] ;B[ci] ;W[cj] ;B[bi] ;W[dj] ;B[bh] ;W[ml] ;B[mm] ;W[lm] ;B[nl] ;W[nk] ;B[ln] ;W[ll] ;B[kn] ;W[mn] ;B[mo] ;W[rm] ;B[rl] ;W[ro] ;B[sn] ;W[pn] ;B[oo] ;W[op] ;B[no] ;W[sm] ;B[pp] ;W[nc] ;B[nb] ;W[ob] ;B[pb] ;W[od] ;B[pc] ;W[mb] ;B[oa] ;W[mc] ;PM[2]MN[1]FG[259:]C[Positional judgement of this position. Any territory uses the concept "current territory". Preliminary prisoner difference = 1] ;B[tt]C[White can play this forcing sequence with suitable timings. The positional judgement after the sequence determines the positional judgement before the sequence.] ;W[rk] ;B[sl] ;W[so] ;B[rn] ;W[pk] ;B[qm] ;W[sn] ;B[on] ;W[qk]C[representative move] ;B[rn] ;L[ec][pe]PM[2]MA[as][bs][cs][cr][dr][ds][er][es][fr][fs][ah][ag][bg][cg][ch][bf][af][ae][ca][da][db][dc][eb][ea][fa][fb][ob][pa][qa][qb][qc][rc][rb][ra][sa][sb][sc][mn][mp][mq][nn][np][nr][ns][op][oq][or][os][pm][pn][po][qn][rm][sm][sn][so][ro][sp][rp][sq][rq][qq][pr][qr][rr][sr][ps][qs][rs][ss][ga][ha][ia][ja][ka][la][nj][nd][ne][nf][ng][of][og][oh][oi][oj][pg][ph][pi][pj][qg][qh][qi][rh][ri][rj][sh][si]C[Follow-up position after the representative forcing sequence. Sequence's prisoner difference = 1. Initial prisoner difference = preliminary prisoner difference + sequence's prisoner difference = 2. The intial prisoner difference is used for positional judgement of this follow-up position. A ~= -1/2 B = -1 Taking into account the initial prisoner difference 2 and the komi 7.5. According to chapter 4.17, the territory count may approximate the area count in this game played with area scoring rules. Territory count = (9 + 11 + 40) - (37,5 + 10) + 2 - 7,5 = 60 - 47,5 + 2 - 7,5 = 7 points in B's favour.]FG[259:] ( ;W[tt] ;B[tt] ; ) ( ;PM[2]C[Start of a reduction sequence determining the territory in the lower left.]FG[259:] ;W[tt] ;B[cp] ;W[cq] ;B[co] ;W[dq]C[This move is the most peaceful reply when defending the territory. Local territory count = -10. (Negative values favour White.)]MA[as][bs][cs][cr][dr][ds][er][es][fr][fs] ) ( ;PM[2]FG[259:] ;W[jq]L[lq]C[If we assume this middle game move, then W does not have any follow-up sente reduction, e.g., A could not be understood as sente. Therfore it is unhelpful to assume the move as the expected middle game move for the sake of reducing the lower right territory.] ) ( ;PM[2]C[Determination of the lower right territory.]FG[259:] ;W[iq]C[Some such move can be expected during the middle game if W plays first on the lower side.] ;B[tt] ( ;W[kq]C[This is the farthest W can extend to reduce in sente. Any farther extension could not be understood well as a sente.] ( ;PM[2]FG[259:] ;B[lq] ;W[kr]C[The sente move.] ;B[lr] ;W[ls] ;B[ms] ;W[ks] ;B[mr] ;W[kp] ;B[lp] ;W[ko] ;B[lo] ;PM[2]MA[ss][sr][rr][rs][qr][qs][pr][ps][or][os][nr][ns][mq][mp][np][oq][op][qq][rq][rp][sp][sq][so][ro][sn][sm][rm][qn][pm][pn][po][nn][mn]C[Local territory count = 40. Note that the earlier prisoners are accounted globally.]FG[259:] ) ( ;B[lr]C[B keeps less.] ) ) ( ;PM[2]FG[259:] ;W[kr] ;B[mr]C[Up to here, it can be called W's sente.] ;W[tt] ;B[kq]C[B has this follow-up forcing sequence.] ;W[jq] ;B[lr] ;W[kp] ;B[lq] ;W[ks] ;B[ls] ;W[jr]L[jp]C[B has less territory than in the standard sente sequence but the weakness A is a greater disadvantage for W. Therefore, usually W does not choose the first move of this variation.] ) ) ( ;PM[2]C[Determination of the upper right boundary of the W moyo and assuming that B cannot threaten to connect to T8 because R9 might be at T9.]FG[259:] ;W[tt] ( ;B[sf]C[Small monkey. This can be understood as sente.] ;W[sg] ;B[se] ;W[rg] ;B[sk] ;W[sj] ) ( ;PM[2]FG[259:] ;B[sg]C[big monkey] ;W[rg] ;B[se] ;W[sh] ;B[sf]C[Not necessarily sente, especially if R9 is at T9. Therefore, the big monkey is not chosen for the determination of the upper right moyo boundary.] ) ) ( ;PM[2]C[Determination of a center boundary of the territory of the W moyo stretching from the upper to the right side.]FG[259:] ;W[tt] ;B[gd]C[Since peaceful defense is assum
Re: [Computer-go] Final score 4-1 - congratulations to AlphaGo team!
Congratulations! On 15.03.2016 13:10, Petr Baudis wrote: AlphaGo has won the final game, tenaciously catching up after a tesuji mistake in the beginning No. Do not trust Redmond's positional judgement. IMO, after the initial tesuji sequence, the position was balanced. (Kim said: White was a little better at that moment.) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo & DCNN: Handling long-range dependency
On 14.03.2016 08:59, Jim O'Flaherty wrote: an AI player who becomes a better and better teacher. But you are aware that becoming a stronger AI player does not equal becoming a stronger teacher? Teachers also need to (translate to and) convey human knowledge and reasoning, and adapt to the specific pupils' needs (incl. reasoning, subconscious thinking and psychology) while interacting with human language specialised in go language. Solve two dozen AI tasks, combine them and then, maybe, you get the equivalent of a teacher. [FYI, I have taught 100+ regular single go pupils since 2008, and groups of pupils.] -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo & DCNN: Handling long-range dependency
On 14.03.2016 09:33, Petri Pitkanen wrote: And being 600 elo points above best human you are pretty close to best possible play. You do not have any evidence for such a limit. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Game 4: a rare insight
On 14.03.2016 03:17, Horace Ho wrote: According this analysis, move 78 is not a "miracle" move ... http://card.weibo.com/article/h5/s#cid=23041853a2e03d0102w6rl; I have not had time to verify the tactics by reading yet but suppose this webpage's sequences are right, move 78 and the preceding sequence is a well-timed, cute trick play and the Alphago teams needs to understand why the trick worked. I'd guess that it would have found a correct reply if the moyo defense had been a local, short term problem. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] Verification Reading for Probabilities
Suppose an MC/NN program suggests a move with the supposedly highest winning probability. Due to the imperfect information for suggesting this move, I suggest to apply my human player principle "verify by reading" by verifying the suggested move by reading. This can use the following method "verification reading for probabilities", which I suggest now: We have the current position with a winning probability and the suggested first move in the current position. Do _reading_ to quiet leaves, for which we know or determine a winning probability. A decision in the tree is a success iff it results in _at least_ the same probability [alternative: a threshold probability] as that of the current position. I.e., the verification reading shall ensure that the winning probability does not decrease as a consequence of starting with the suggested move. Such should prevent the program from suddenly experiencing harshly dropping probabilities during a few successive moves / mistakes of the game, such as in game 4 of the AlphaGo - Lee Sedol match. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Congratulations to AlphaGo
On 12.03.2016 22:03, Thomas Wolf wrote: We currently have no measure at all to judge how safe a winor loss is at any stage of the game. We have my theory according to my books for assessing the territorial and the dynamic aspects (development directions, neutral stones, statuses (incl. those of potential invasion groups), options, aji, invasions, reductions, (local) potential, influence, thickness, fights) of every position. The theory does not provide a single number (such as a one-dimensional probability) but judgement need not be one-dimensional and can depend on reading to assess particular aspects, such as a status. The only valid strength indicator would be to gradually increase handicap stones or komi for the previous loser in a series of games. Altering komi is much better than altering handicap because komi can be adjusted in finer steps and does not artificially restrict strategy a lot. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo & DCNN: Handling long-range dependency
On 11.03.2016 08:24, Huazuo Gao wrote: Points at the center of the board indeed depends on the full board, but points near the edge does not. I have been wondering why AlphaGo could improve a lot between the Fan Hui and Lee Sedol matches incl. learning sente and showing greater signs of more global, more long-term planning. A rumour so far suggests to have used the time for more learning, but I'd be surprised if this should have sufficed. So far, I have the following theories: - deeper net - greater parameters for convolutional patterns (instead of 5x5 and 3x3, (also) use larger parameters) or combine the earlier parameters with additional larger parameters or with an additional NN having only / mostly larger parameters - replace or enhance top KGS games by 100,000+ pro games - instead of / in addition to feed forward nets, use long short term memory nets (but I cannot know if this is advantageous considering presumably greater GPU time) - instead of single position patterns, use combinations of current position and later positions, for different (dynamic) parameters of time shift, so as to model long-term effects -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Finding Alphago's Weaknesses
On 10.03.2016 16:48, Darren Cook wrote: in game 2, black 43 and 45 were described as "a little heavy". It did seem (to my weak eyes) to turn out poorly. I'm curious if this was a real mistake by AlphaGo, or if it was already happy it was leading, and this was the one it felt led to the safest win? In human terms, it was a combination of: limitation of the expansion potential of the white left side, shinogi, sente and developing the potential of the upper side including its center potential. Ugly and marvellous strategy of simplifying the game (same: reduction of the right side in sente) and creating a winning position by robbing White of every option of creating significant new territory regions / expansions. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] Finding Alphago's Weaknesses
On 10.03.2016 00:45, Hideki Kato wrote: such as solving complex semeai's and double-ko's, aren't solved yet. To find out Alphago's weaknesses, there can be, in particular, - this match - careful analysis of its games - Alphago playing on artificial problem positions incl. complex kos, complex ko fights, complex sekis, complex semeais, complex endgames, multiple connection problems, complex life and death problems (such as Igo Hatsu Yoron 120) etc., and then theoretical analysis of such play - semantic verification of the program code and interface - theoretical study of the used theory and the generated dynamic data (structures) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] Positional Judgements on AlphaGo / Fan Hui Games
I have written commentaries on positional judgements of the games, see usenet <176jdb1f33rlnlm1nnrmo1g3g5nme16...@4ax.com> <vrhndb1juu1dof00dmdo6pukgrl0m2c...@4ax.com> or web http://www.lifein19x19.com/forum/viewtopic.php?f=17=12766 http://www.lifein19x19.com/forum/viewtopic.php?f=15=12771 alternative http://www.dgob.de/yabbse/index.php?topic=5977.0 http://www.dgob.de/yabbse/index.php?topic=5978.0 *** Conclusion: AlphaGo's implicit positional judgement is on average better than the judgement of weak professionals. It is an open question whether AlphaGo judges huge spheres of dominating influence correctly because its skill in reducing them as at least strong amateur dan level, but probably pro level. Implicitly, the program complies with the best available explicit human theory when reducing a big moyo. This includes model short-term use of aji in all the moyo's boundaries. Possibly with the exception of the huge sphere of dominating influence, AlphaGo also makes good move choices when it must play an influence stone and greatly alter the influence balance. IMO, Lee needs to take advantage of whole board, long-term interaction or take sufficient advantage of the program's revealed strategic mistakes to demonstrate AlphaGo's limits. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Move evalution by expected value, as product of expected winrate and expected points?
On 23.02.2016 11:36, Michael Markefka wrote: whether one could train a DCNN for expected territory First, some definition of territory must be chosen or stated. Second, you must decide if territory according to this definition can be determined by a neural net meaningfully at all. Third, if yes, do it. Note that there are very different definitions of territory. The most suitable definition for positional judgement (see Positional Judgement 1 - Territory) is sophisticated and requires a combination of expert rules (specifying for what to detemine, and how to read to determine it) and reading. A weak definition could predict whether a particular intersections will be territory in the game end's scoring position. Such can be fast for MC or NN, and maybe such is good enough as a very rough approximation for programs. For humans, such is very bad because it neglects different degrees of safety of (potential) territory and the strategic concepts of sacrifice and exchange. I have also suggested other definitions, but IMO they are less attractive for NN. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Congratulations to Zen!
Aja, sorry to bother you with trivialities, but how does Alphago avoid power or network failures and such incidents? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] Knowledge Details
The current fashion favours general AI approaches forgoing knowledge details. Given enough calculation power applied to well chosen AI techiques, many knowledge details are redundant because they are generated automatically: AlphaGo does play (at least some) ko fights with ko threats, tesujis, test moves, (at least some) life and death or semeai problems etc. At the same time, AI calculation power is still not large enough to generate all human knowledge details. Aji with long-term impact and maintaining the life status "independently alive" instead of unnecessarily transforming it to "(ko|independently alive)" (aka "unsettled") are prime examples. Programs also play for the win regardless of whether moves are suboptimal for the score difference - human players tend to avoid such (programs would also profit from avoiding such to prevent losing when making a later mistake due to a knowledge gap related to insufficient error handling). There is another great threat related to knowledge details, which is not immediately apparent and will be even much less apparent when programs will exceed top human playing strength: A program can run into a situation where an infrequent knowledge detail becomes relevant. And a program can run into ordinary software or hardware bugs, something that must be detected and correct on the AI level. My conclusion is: human expert knowledge on details of go theory matters. There have been 9p players committing self-atari when filling a dame, so you might argue that programs may infrequently make similar blunders. When I issued a million dollar prize, I'd prefer human expert knowledge implemented at least as an additional layer of error handling. (Other fun includes internet connection trouble, server bugs of distributed computers, hardware bugs of the local interface computers or interrupted power supply.) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] 57%
AlphaGo is said to predict 57% of professionals' moves. How is this number measured and from which sample? At some turns, there is only one correct move - at other turns, strong go players would say that there are several valid supposedly correct moves. This is one of the reasons why 100% cannot be the optimum but a smaller percentage must be the best. Pro players, or players of the database sample (incl. real world 3d players being 9d on KGS), make mistakes. A neural net learns from a sample and therefore also learns the mistakes. This is the most important reason why 100% cannot be the optimum but a smaller percentage must be the best. (Roughly) which percentage is optimal? Why? Is the optimum greater or smaller than 57%? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Knowledge Details
On 03.02.2016 15:34, Jim O'Flaherty wrote: BTW, I have my own personal aspirations which have been thwarted by this development. I have several thousand hours of doing my own research and development [...] although I will likely drift further away from Go as the focal point of motivation. Maybe I should know but I have not always closely followed the relations between persons and computer project / research names. Therefore please let me ask: Which have been your personal aspirations, motivations and research investments? Best of luck finding your way through your meaning and value (emotional) reintegration of this newest reality update. Nothing has changed (or will change when "brute force" surpasses top human play) for me because my main research goals are the strong solution of go under every go ruleset and the explanation of go theory to human players (incl. myself). -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Knowledge Details
On 04.02.2016 02:52, David Ongaro wrote: At the same time I've to point out that you seem to plan to get very old. I will not see the solution, which needs at least another 400 years unless computers learn to research. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks and Tree Search
On 02.02.2016 17:29, Jim O'Flaherty wrote: AI Software Engineers: Robert, please stop asking our AI for explanations. We don't want to distract it with limited human understanding. And we don't want the Herculean task of coding up that extremely frail and error prone bridge. Currently I do not ask a specific AI engine explanations. If an AI program only has the goal of playing strong, then - while it is playing or preparing play - it should not be disturbed with extra tasks. Explanations can come from AI programs, their programmers, researchers providing the theory applied in those programs, researchers analysing the program codes, data structures or outputs. I do not expect everybody to be interested in explanations, but I ask those interested. It must be possible to study theory for playing programs, their data structures or outputs and find connections to explanatory theory - as much as it must be possible to use explanatory theory to improve "brute force" programs. Herculean task? Likely. The research in explanatory theory is, too. Error prone? I disagree. Errors are not created due to volume of a task but due to carelessness or missing study of semantic conflicts. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mathematics in the world
On 02.02.2016 13:05, "Ingo Althöfer" wrote: when a student starts studying Mathematics (s)he learns in the first two semesters that everything has to be defined waterproof. Later, in particular when (s)he comes near to doing own research, you have to make compromises - otherwise you will never make much progress. When I studied maths and theoretical informatics at FU Berlin (and a bit at TU Berlin) (until quitting because of studying too much go, of course), during all semesters with every paper, lecture, homework or professor, everything had to be well-defined, assumptions complete and mandatory proofs accurate. As a hobby go theory / go rules theory researcher, I can afford the luxury of choosing formality (see Cycle Law), semi-formality (see Ko) or informality (in informal texts) because I need not pass university degrees with the work. My luxury of laziness / convenience when I use semi-formal style (as typical in the theory parts of my go theory papers) indeed has the advantages of being understood more easily from the go player's (also my own) perspective and allowing my faster research progress. If I had had to use formal style for every text, I might have finished only half of the papers. If we can believe Penrose (The Road to Reality) and Smolin (The Trouble with Physics), the world of mathematical physics is split into guesswork (string theory without valid mathematical foundation) and accurate maths. Progress might not be made because too many have lost themselves in the black hole of ambiguous string theory. Computer go theory seems to be similar to physics. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks and Tree Search
On 02.02.2016 20:21, Olivier Teytaud wrote: On the other hand, they have super strong people in the team (at the pro level, maybe ? if Aja has pro level...) Ca. 5d amateur in the team is enough, regardless of whether Myongwan Kim thinks that only 9p can understand. Not so. Kim's above 5d amateur comments were related to reading or by heart knowledge of the latest nadare variations (before the post-joseki aji mistakes, which can be detected by 5d, or even below), but reading / joseki is not AlphaGo's weakness. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks and Tree Search
On 02.02.2016 19:07, David Fotland wrote: consider some of this as the difference between math and engineering. Math desires rigor. Engineering desires working solutions. When an engineering solution is being described, you shouldn't expect the same level of rigor as in a mathematical proof. Often all we can say is something like, "I tried a bunch of things, and this one worked best". Both have value. Of course. This is perfectly fine. - I have criticised something else: the hiding of ambiguity of things portrayed as maths when statements of the kind "this is a heuristic / engineering / first guess" are easily possible. Research papers should be honest. (They may hide secret details, but this is another topic.) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks and Tree Search
On 01.02.2016 23:01, Brian Cloutier wrote:> I had to search a lot of papers on MCTS which > mentioned "terminal states" before finding one which defined them. > [...] they defined it as a position where there are no more legal > moves. On 01.02.2016 23:15, Brian Sheppard wrote: You play until neither player wishes to make a move. The players > are willing to move on any point that is not self-atari, and they are willing to make self-atari plays if capture would result in a Nakade (http://senseis.xmp.net/?Nakade) Defining "terminal state" as no more legal moves is probably inappropriate. The phrase "willing to move" is undefined, unless they exactly define it as "to make self-atari plays iff capture would result in a Nakade". This requires a proof that this is the only exception. Where is that proof? It also requires a definition of nakade. Where is that definition? In my book Capturing Races 1, I have outlined a definition of "[semeai-]eye" and, in Life and Death Problems 1, of "nakade". Such are more complicated by far than naive descriptions online suggest. In particular, such outlined definitions depend on the still undefined "essential [string]", "seki" [sic, undefined as a strategic object because the Japanese 2003 Rules' definition does not distinguish good from bad strategy!] and "lake" [connected part of the potential eyespace..., which in turn is still undefined as a strategic object]. They also depend on "ko", but at least this I have defined: http://home.snafu.de/jasiek/ko.pdf Needless to say, determining the objects that are essential, seki, lake, ko is a hard task in itself. So where is the mathematically strict "definition" of nakade? Has anybody proceeded beyond my definition attempts? I suspect the standard problem of research again: definition by reference to a different paper with an ambiguous description. If ambiguous terms are presumed for pragmatic reasons, this must be stated! My mentioned terms are ambiguous but less so than every other attempt - or where are the better attempts? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mathematics in the world
On 02.02.2016 13:05, "Ingo Althöfer" wrote: For research in general it is good to have waves: Research is faster if informalism and formalism progress simultaneously (by different people or in different papers). -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks and Tree Search
On 02.02.2016 11:49, Petr Baudis wrote: you seem to come off as perhaps a little too aggressive in your recent few emails... If I were not aggressively critical about inappropriate ambiguity, it would continue for further decades. Papers containing mathematical contents must clarify when something whose use or annotation looks mathematical is not a definition / well-defined term but intentionally ambiguous. This clarity is a fundamental of mathematical, informatical or scientific research. Without clarity, progress is delayed. Every professor at university will confirm this to you. The question was about the practical implementation of an MC simulation, which does *not* require formal definitions of all concepts used in the description, or any proofs. It's just a heuristic, and it can be arbitrarily complicated, making a tradeoff between speed and accuracy. Fine, provided it is clearly stated that it is an ambiguous heuristic and not an [unambiguous] definition / term. References / links (possibly iterative) hiding ambiguity without declaring it are inappropriate. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks and Tree Search
On 01.02.2016 15:15, Jim O'Flaherty wrote: I'm not seeing the ROI in attempting to map human idiosyncratic linguistic systems to/into a Go engine. Which language would be the one to use; English, Chinese, Japanese, etc? As abstraction goes deeper, the nuance of each human language diverges from the others (due to the way the human brain is just a fractal based analogy making engine). [...] > unless you are, of course, suggesting that is something you are taking up. :) The human language for interaction with / translation to programming language includes - well-defined terms / concepts - rules / principles with stated presuppositions - methods / procedures / informal algorithms - proofs / strong evidence for the aforementioned being correct / successful (always or to some extent) Of course, I am an example of a person having been doing this for many years. In fact, I might be the leading generalist for go theory expert knowledge stated in writing. The AI world is changing to make explaining computation cognition to humans less necessary, or even desirable. I disagree strongly. Almost all the AI world has done is creating strong programs. Explaining human thinking and explaining program thinking in terms of human thinking is as important as it has always been. Why bound the solution space to only what cognitively linguistically limited humans can imagine and/or consider? Indeed. I prefer to exceed limitations by creating new terms, definitions for undefined terms, principles, methods etc. Human beings can better learn if they know what to learn because the contents is described clearly. about what is rapidly approaching as human cognition automateable. Eh? Besides GoTools, there has been very little, AFAIK. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks and Tree Search
On 01.02.2016 14:38, Aja Huang wrote: AlphaGo may do much better in tactical situations than Crazy Stone and Zen. Judging very quickly from the Fan Hui games, AlphaGo's group-local "reading" is very deep and accurate but I'd need to read for myself equally deeply and carefully before I would want to confirm Myongwan Kim's related opinion. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo and the Standard Mistake in Research and Journalism
On 31.01.2016 17:19, John Tromp wrote: It will never be known since there's not enough space in the known universe to write it down. We're talking about a number with over 10^100 digits. How do you know that an implicit expression (of length smaller than 10^80) of the number does not exist? :) > [interesting stuff deleted] It is reasonable to expect the perfect komi does not depend on games of more than 361 moves. I do not think we may make such a premature claim. Even with some ko fights, the ko recaptures are likely bounded by the number of unplayed points. From experience as go players, yes. But... I have seen too many surprising sequences and sacrifices to be sure. Then we estimate the decision complexity to be upper bounded by 200^181 and the game tree complexity by 200^361. I won't believe such until proven:) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks and Tree Search
On 31.01.2016 20:28, Peter Drake wrote: pick a new research topic. - explain by the program to human players why MC / DNN play is good in terms of human understanding of the game - incorporate the difficult parts, such as long-term aji - solve the game: prove the correct score, prove a weak solution, prove a strong solution [These mathematics keep us busy for at least 400 years unless bot research occurs earlier.] - create computers that act as mathematicians incl. creativity, invention of propositions and their proving [so that the bot researchers can solve the game faster] - teach the computer expert knowledge so that a) MC / DNN bots become even stronger and b) programs can teach with explanation and reasoning understood by human pupils - apply computer go research to other fields while ensuring that the humans cannot be the victims of bugs and ambiguous responsibilty towards law and ethics [medicin or cars: who goes to jail if AI kills people, how to prevent AI from ruling the world] - Play "Conway / Jasiek": modify the rules, invent new games, apply computers. Enough for research for centuries if not millenia, I'd say. "Game over / intelligence solved" - never heard greater nonsense before. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] AlphaGo and the Standard Mistake in Research and Journalism
On 31.01.2016 19:57, John Tromp wrote: What is your best estimate of point where where chances are even? I do not know. what numbers the press could use that are not too arbitrary. - The number P of legal positions. - An empirical average number I of available intersections for the next move, where I needs to be determined before exposed to the press. - The maximal number N of available intersections for the next move. - An empirical average number X of moves in not resigned games, where X needs to be determined before exposed to the press. - The maximal number ca. 450 of moves in a practically occurring game [excluding neutral intersections, area scoring removals and virtual pass fights]. - The maximal number ca. P of moves in theory. - The number of legal games is unknown. An explanation that at every move, all currently available, legal intersections must be considered and the upper bound is ca. N^P. For the yellow press: "The number of 10^80 atoms in the universe is much smaller than the number 2 * 10^170 of possible positions, which is very much smaller than the uncountable number of possible different games." -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
[Computer-go] AlphaGo and the Standard Mistake in Research and Journalism
According to John Tromp et al at http://tromp.github.io/go/legal.html the number of legal 19x19 go positions is P19 = 2081681993819799846 9947863334486277028 6522453884530548425 6394568209274196127 3801537852564845169 8519643907259916015 6281285460898883144 2712971531931755773 6620397247064840935 P19 ~ 2.081681994 * 10^170 *** The number G19 of legal games under a given go ruleset is unknown. We only know lower bounds and the following upper bound, where N = 19x19 = 361 is the number of intersections on the board: For positional superko (prohibition of recreation of the same position after completion of a move on the board), no passes, and no resignation, the number of possible games is smaller than N^P19 because P19 also restricts the maximal number of moves per game and there are at most N possible intersections per move. So approximately the upper bound for the number of legal games under the mentioned rules is G19 < 361^(2.081681994 * 10^170) Note that, changing the ko rules from positional superko to Japanese / Korean / Chinese style no-result ko rules, we get an infinite number of legal games with equivalence classes for as many different games as under positional superko. (For the sake of simplicity, let us ignore the impact of passes and resignations as technicalities for now.) *** Citation from my rec.games.go Rules FAQ: "Extremely modest estimates look like 10^N, which is based on an assumption of 10 reasonable intersections per move. A popular estimate for 19x19 go is 10^761, which must have originated from a typo and should be 10^361, if at all... These types of estimates are so popular because humans cannot even imagine the suggested number of atoms in the universe, 10^80. For comparison, 3^361 is ca. 10^172." In the AlphaGo research paper https://storage.googleapis.com/deepmind-data/assets/papers/deepmind-mastering-go.pdf the number of possible go games is estimated as approximately b^d, where b is the breadth (number of available intersections per move) and d is the depth (length of a game as its number of moves) and stated as approximately 250^150 by referencing to an earlier research article. This number is not completely meaningless: it is somehow related to empirical data. Even so, the number would be wrong because practically occurring games have been reported to last up to ca. 450 moves. Supposing 250 would be a reasonable average number of available intersections, we would get 250^450. I.e., already by correcting a minor mistake in the referenced number, we get a more meaningful number that already is astronomically larger than the stated 250^150. If we pretend the 250 in 250^450 to be reasonably correct, this more correct empirical number serves as an empirical lower bound. *** Therefore, under the made assumptions (such as permitting an empirical estimate for the lower bound), we know that the number G19 of legal go games is 250^450 < G19 < 361^(2.081681994 * 10^170). *** In various go blog messages and media articles, the same kinds of wrong (by astronomic factors too small) numbers are circulated for "the" complexity of go: 10^361, 10^170, 10^171, 250^150. The AlphaGo research paper, go players feeding journalists with information and journalists make the same mistake: they copy or cite without understanding the meaning of a particular number. The worst example has been a popular science journal's statement that ca. 10^170 would be the number of possible go games (when confusing it with the extremely smaller number of possible go positions). This leads us to types of complexities. There are the complexity of the number of legal positions, the complexity of the number of legal games (which is the actual complexity of go as a game because great applicable shorthands to a solution of perfect play are unknown) and the computational complexity of the generalised game on boards with N intersections (but go as a game is played on the 19x19 board with its fixed N = 361). -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Game Over
On 28.01.2016 04:57, Anders Kierulf wrote: Please let me know if I misinterpreted anything. You write "Position evaluation has not worked well for Go in the past" but I think you should write "...Computer Go..." because applicable, reasonably accurate theory for human players' positional evaluation exists, see e.g. my two books Positional Judgement. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Mastering the Game of Go with Deep Neural Networks and Tree Search
Congratulations to the researchers! On 27.01.2016 21:10, Michael Markefka wrote: I really do hope that this also turns into a good analysis and teaching tool for human player. That would be a fantastic benefit from this advancement in computer Go. The programs successful as computer players mostly rely on computation power for learning and decision-making. This can be used for teaching tools that do not need to provide text explanations and other reasoning to the human pupils: computer game opponent, life and death playing opponent, empirical winning percentages of patterns etc. Currently such programs do not provide sophisticated explanations and reasoning about tactical decision-making, strategy and positional judgement fitting human players' / pupils' conceptual thinking. If always correct teaching is not the aim (but if a computer teacher may err as much as a human teacher errs), in principle it should be possible to combine the successful means of using computation power with the reasonably accurate human descriptions of sophisticated explanations and reasoning. This requires implementation of expert system knowledge adapted from the best (the least ambiguous, the most often correct / applicable) descriptions of human-understandable go theory and further research in the latter. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Number of Go positions computed at last (John Tromp)
On 25.01.2016 09:11, Mark Goldfain wrote: I haven't heard of anyone else doing similarly interesting work on the theoretical foundations of the game. Sorry, but this is your fault. Where Tromp excels is Go combinatorics, a field that does not equate the more general "the theoretical foundations of the game". For the latter (and excluding computer go research), there have been a few researchers with more contributions, even if you permit formal theory only and exclude more general math research (e.g. by John Conway) also applicable to go. > One of the most frustrating things about writing a program to play go > is that the rules are a bit blurry. You say so after having read my commentaries? http://home.snafu.de/jasiek/rules.html On 25.01.2016 10:15, Olivier Teytaud wrote: > the Chinese rules Please do not confuse the flawed Chinese Rules http://home.snafu.de/jasiek/c2002.pdf http://home.snafu.de/jasiek/c2002com.pdf with the flawless Area Scoring in general. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Number of Go positions computed at last
On 22.01.2016 05:18, John Tromp wrote: It's been a long journey, and now it's finally complete! http://tromp.github.io/go/legal.html Congratulations! You must have needed 15 or 20 years of research to find the result? Eventually you heavily rely on computational power. How has it been possible to get hold of the computers and computation time? When described in informal words, how have you attacked and proceeded with the theory of the problem? What can other researchers learn from your experience of how to research well? The number of legal positions itself seems like a piece of trivia (is it?) but why do you think it is important to have determined the number, that is, what is the research benefit? If I may ask, what has been your motivation beyond curiosity? You mention the calculation to be a server benchmark. Have there been other equally or more suitable server benchmarks or is this particular problem ground-breaking as a server benchmark? What do the solution and its theory tell go players for tactics and strategy and go programmers for developing better go playing programs? Does the solution give a useful clue of how difficult it is to solve go as a game weakly or strongly? That is, how is the number of legal positions related to the computational complexity in time and space of solving the 19x19 go game (under a given go ruleset) if viewed as the specific 19x19 problem and not as the context of the general nxn problem's class of computational complexity? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Seki frequencies
On 18.01.2016 15:58, "Ingo Althöfer" wrote: understand the development of the fight during the running game. Simply speaking, there is a relation between a) captures / avoiding captures and thus controlled regions of the board and b) score peaks. This is so for ordinary captures, ordinary unsettled life / death, maintained / broken territory regions and sekis. Correct plays or mistakes result in regions controlled by certain players. Peaks tell you where such position-local changes in scores occur and you can call these regions "potential fighting regions". Once you know where they are, a second step might do a functionally abbreviated reading (To defend or not to defend this region / string?) for these big decisions together to find potentially good strategic decision-making for the "large" scale decisions. This is not the same as my go player term "fighting region" (simply speaking, a local set of unsettled groups with its near empty surrounding intersections, details see the book Fighting Fundamentals), and it is also not the same as what a go player understands as a "fight", which includes such things as "attack a group for the same of making an additional 1 point in sente during this particular sequence". With histograms, you see nothing like this in the near future, I'd guess. Define 'fight'! (As you understand it for your study purpose.) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Seki frequencies
On 18.01.2016 08:47, "Ingo Althöfer" wrote: My main motivation for thinking about Seki was/is the question if it is possible to recognize upcoming Seki situations in the histograms of an MCTS bot As you have understood now, seki detection is not only a shape / topology question but mainly is about stability due to disadvantageous value changes both locally and globally (e.g. ko threat played in a seki as a means of offering a local sacrifice). In terms of histograms, you might identify the number of local value extrema e.g. for the whole board position's score. Is seki much different from ordinary life and death? Why care if anything is a seki? Maybe it is sufficient to study histograms of final global scores and detect at which tree branches peaks arise. (Or do likewise for subpositions.) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Seki frequencies
On 17.01.2016 12:19, "Ingo Althöfer" wrote: Can you give an example for anti-seki? One black string and one white string share exactly one liberty and do not have any other liberty. Copy & paste the same but invert colours elsewhere on the board. The shortest perfect play is to pass. It can also be more complicated: use pendulum kos. Asymmetric sets of subpositions is an interesting study field for algebraists, says Charles Matthews. Listing the possible configurations is a demanding open research field. Perhaps you and someone like Thomas Wolf (with his life-and-dath background) would be "the right" people for this question. Actually for non-ko sekis there are specialists such as Bill Taylor, Harry Fearnley, Ger Hungerink, Denis Feldmann and others whose name I do not recall now. Given enough time, Wolf and I might also do related research and Wolf would have suitable programs to use but I have - for me - more urgent tasks, such as continuing research on ordinary semeais. Would strong go bots also fall in this category? I do not know because I have seen too few games played by them. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Seki frequencies
On 17.01.2016 05:33, "Ingo Althöfer" wrote: Seki means a constellation on the go board with two living neighboring groups: one by Black, the other one by White. Each of the groups has only one eye. And they share a joint liberty. Seki has AT LEAST two groups. Sekis can have various different shapes incl. such with unreducable three shared inside liberties per pair of groups, major strings having only 1 liberty (hane seki), or multiple kos. Do not overlook the stable anti-sekis (stable because other anti-sekis exist elsewhere on the board). Listing the possible configurations is a demanding open research field. My question: How frequent are Seki constellations? This very greatly depends on which player population is observed. On KGS, sekis are frequent. Among Japanese professionals (for which I counted sekis for an unrepresentative sample from the second half of the 20th century), sekis occur only once in ca. every 70th game. I think sekis are not so scarce among Chinese and Korean professionals. Apparently long playing time combined with great playing strength avoids sekis. Short thinking time with relatively great playing strength (amateur dans on KGS) seeks seki as a reasonable compromise in unreasonable fights. *** If you want to study seki configurations, the best is to do theoretical research. If you want many ordinary sekis in actual games, take KGS samples. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Go Aesthetics
Is playing bad moves good for aesthetics? No? Then why call it aesthetics? Call it perfect / good play. The most "beautiful" stone is bad if it is dead. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] History of influence concepts in go
On 10.01.2016 20:06, "Ingo Althöfer" wrote: it seems that different people are using the name "influence" for different objects/properties. In the computer go scene, 1-dimensional use goes back to Albert Zobrist in his doctoral dissetation from 1970. Where does your framework for "multi-dimensional" influence comes from? Influence is a well-known term among players but its clear meaning remained a mystery until I described in 2011. Before influence was so unclear that it was hard, as a player, to know its difference from the other related term thickness. Early player understanding of influence was as naive as black / white influence light decreasing by distance (i.e., it is not really influence but proximity multiplied by some radiation function such as 1 or 1/x^2 or Manhattan distance and negative for white light and maybe visually blocked by the disks of stones) but everybody knew that that was wrong because dead stones do not create as much influence as live stones. Early expert system programs used the same naive concept, and every programmer would use his own implementation of distance and intensity of light. Such light maps give colourful maps that are impressive as paintings but close to useless because of containing both correct and false information. Stronger players know that influence and thickness are related to strength of the stones creating the influence and solidity of the groups of thickness. But what is strength? From traditional Asian go theory, it was known that there is some relation between strength and strategic concepts such as (little) aji, development directions, board partitioning lines, potential for further territory etc. However, a systematic assessment of strength was missing. So I studied the fundamentals of the traditional go theory and noticed that several strategic concepts (such as aji) used for thickness and influence were just implications of the more basic strategic concepts of connection, life and territory. I invented / (for 'life') rediscovered degrees of connection, life and territory, distinguished influence (the property of affecting other intersections) from thickness / thick shape stones (the property of the "strong" stones creating the influence) and defined both in terms of degrees of connection, life and territory. Territory is optional in the definition and can also be studied independently. Territory as a propery makes sense because it makes a difference whether influence cannot be used because of being in a neutral region or whether it is / can be used for protecting existing / making additional territory. Study a few simple examples of groups of strong stones with a few or more opposing stones in the neighbourhood, and you notice that degrees of connection and degrees of life can differ from each other. So influence / thickness must be described at least by these two degrees. Furthermore, the values differ for Black and White, so at least four parameters are necessary for a complete description. You find my informal definitions here http://senseis.xmp.net/?Influence http://senseis.xmp.net/?Thickness or more carefully in my books. For the precise parameters of connection and life see http://senseis.xmp.net/?NConnection http://senseis.xmp.net/?NAlive Concepts of proximity should be called 'proximity' while concepts of influence should be called 'influence'. Proximity maps / functions do not explain influence except for the simplest examples in which all stones are alive and the view is clear in every direction. Computer go can have various study purposes (such as training neural nets or predicting the final colour control in a scoring position) and some sort of function over all intersections assigning them a single number may be convenient for fast numerical training, but do not forget that such a simplication trains both correct and false information without distinguishing them properly. If we want to become stronger players or create stronger programs, we must distinguish correct from false information. Therefore, replace 1-dimensional by multi-dimensional values if the task is to assess current positions rather than final scoring positions, in which one value is sufficient. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Using gogui to visualize a CNN for move prediction
On 09.01.2016 22:44, Justin .Gilmer wrote: influence = A1 0.3 What am I missing? Influence is not a one-value property. If the property is something else, please give it a different name. If it shall be influence, describe it by at least four integer values in terms of n-connected and m-alive from either player's perspective, allowing UNDEFINED where necessary. For the sake of simplicity, n, m > 2 or < -2 you can replace by 2 or -2. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Komi 6.5/7.5
On 06.11.2015 10:47, Aja Huang wrote: area scoring, in which case the score is almost always odd. Black wins: odd score White wins: even score -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Komi 6.5/7.5
On 04.11.2015 13:59, Aja Huang wrote: Ke Jie said in his opinion on 19x19 komi 6.5 or 7.5 favors White. Go theoretical considerations (see Joseki 2 - Strategy, chapter 4.4.1) estimate the per move value of the first move as ca. 14 points, so suggest the komi 7. Pro game statistics, with the exception of Japanese 2 day games, suggest that 6.5~7.5 is closer to 50% than 5~6. That seems consistent to MCTS's behavior? i.e. on the empty board, with komi 7.5, Black's win rate is usually between 46% and 48% meaning White is ahead. No, this first of all means that MCT should study 7 more. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] re comments on Life and Death
On 10.09.2015 08:24, David Fotland wrote: I would say rather, that expert systems are dead in Go because many smart and talented people, including professional experts, worked diligently for two decades on this approach and none were able to get stronger than about 5 kyu. This is a strong experimental result, not an opinion. This says nothing about the potential of expert systems when done right. General talent, professional expert system designers or professional players are insufficient. What is needed is a very good understanding of go theory on all topics of go theory as expert system knowledge. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] re comments on Life and Death
On 10.09.2015 10:29, Jim O'Flaherty wrote: Perhaps you could give some more concrete examples of what you have done already; i.e. where you have moved from the messy human linguistic/cognitive "principles" to something much more formal? In my principles (or other theory), the degree of ambuigity varies from formal to ordinary language. Example of a formal formula: dF =? 0 where F is 'fighting liberties' and the formula applies to what I call 'class 1 semeais'. Example of (seemingly) ordinary language in a principle about defending life in a fight: "Maintain connection of a group's important strings." This is not ordinary language though but I use 'connection' and 'important string' as consistent terms in all my books, where the former is defined but the latter is (still) undefined. Consistent use of the same terms and defined concepts everywhere and well chosen definitions for the basic terms remove much of the mess and enable hierarchic design and use of principles etc. For several hundred further examples of definitions and principles, see my books and papers. For my first six of 11 books and earlier papers / messages, see the short overview http://home.snafu.de/jasiek/RobertJasiekGoTheoryResearch.html -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] re comments on Life and Death
On 09.09.2015 07:42, David Fotland wrote: I classify groups instead. Each classification is treated differently when estimating territory, when generating candidate moves, etc. This is reasonable. The territory counts depend on the strength of the nearby groups. Whether this is good depends on how you link strengths to counts. *** Was your influence function like radiated light? Such would have too little meaning. Monte Carlo has a big advantage in that it estimates the probability of winning the game, rather than my old approach of trying to estimate the final score. Whether it is an advantage depends on one's objectives. For an expert system, estimating the score is just one aspect for further application and does not finish the job. (To start with, a positional judgement consists of more than the 'territory count' and group strengths of the current position.) -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] re comments on Life and Death
On 09.09.2015 09:53, Petri Pitkanen wrote: Too many contradicting heurestics The mid-term problem is not mutual contradiction of heuristics because their careful study can remove the contradictions and establish a hierarchy of principles. Only the problem of great number of principles to be coded and maybe of the complexity of time remain. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] re comments on Life and Death
On 09.09.2015 16:45, Jim O'Flaherty wrote: I'm not convinced that it's reducible I am convinced it is,... [...] to [...] a [...] set of principles ...where the principles need some dynamic input, such as reading, when necessary. I don't think it can currently be done for a static Go position It can be done. See Positional Judgement 1 - Territory, and I am writing Vol. 2. The principles therein are written for humans and need translation to mathematical definitions (rather easy for me) and additional work on removing fake contradictions (which arise due to ambiguity in the human-orientated language, or not yet spelled out order of priority). However, the hard part has not even been formulating the principles but the decades of study raising my conceptual insight about fundamentals of go theory to a level where principles just fall out as an extra benefit. Static positions must be understood as being quiet or presuming quiescience sequences (settling fights until the position is quiet). I.e., I am not dogmatic about distinguishing static from dynamic positions. Rather static positions can require some dynamic input / reading. (Not surprisingly; quiescience has been familiar to CG for a long time.) I wish you the best of luck producing the set of principles. Luck is useless. What has helped me was careful analytical study of my own (often methodical) thinking. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] re comments on Life and Death
On 05.09.2015 08:00, David Fotland wrote: Completely agree that connections and group strength estimates are key to strength, and are very difficult to get right. From the POV of humans, I have described connection meaningfully. The remaining problem is the variety of application in principles and higher concepts. Whether group strength is needed at all depends very much on what you mean by it. For connections I used shapes and local tactics Shapes are not needed, unless you want to use them to prune tactics. However, if the tactics verification is slow for standard shapes, I'd say this is its fault. > Connection status was used to collect stones into groups. Fine, provided this is not a static partition. Other considerations (such as sacrifice) can make it necessary to alter groups dynamically. For group strength I had about 20 classes with separate evaluators (two clear eyes, one big eyes, seki, semeai, run-or-live, one-eye-ko-threat-to-live, dead-if-move-first, etc, etc). Was group strength an object of several parameters or was it a single number derived from all those parameters? IMO, a single number cannot be meaningful in general. Groups strength was the core concept feeding into the full board evaluation, which tried to estimate the score. But what WAS your group strength...?:) Score estimation of a given position should also depend on territory counts, not only on group strength etc. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] re comments on Life and Death
On 04.09.2015 16:54, Minjae Kim wrote: Why not implement your ideas as a computer program? - I lack time. - Developing my ideas has consumed decades. - I know that there are gaps in my ideas that I need to research in when I will have the additional time: some ideas are formulated for humans but lack formal precision (easy for me but it does require a lot of time for the many ideas); some fields of go theory I could barely explore yet, although I have conceptual ideas of where / how to explore them. - With proceeding research and study, I find more generally applicable ideas replacing some earlier, weaker ideas. In order to keep up with these changes in insight, I'd need much more time for implementation. So realistically I see myself as the author describing ideas useful for both human players and expert systems but others need to implement them and derive their interconnection (such as dissolving seemingly contradicting principles). During later years, I will write more about the latter. Maybe you underestimate the volume of my generated knowledge. Currently, it is (very roughly) 1000 principles, 100 methods, 100 concepts. Maybe it can be compressed to 100, 10, 10 for the sake of expert system input, but even then the implementation task is huge (man-years). Not to mention semantic testing of processed data to get a "thinking" workload similar to my own thinking. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] re comments on Life and Death
On 04.09.2015 17:55, Stefan Kaitschick wrote: It is just too far removed from MC concepts to be productively integrated into an MC system. And no matter what, MC has to be the starting point No. It is also possible to construct it the other way round. Start with an expert system. Whenever that needs some "calculation" and basic counting or limited reading fail, MC can come in to do the calculation. E.g., an expert system can identify groups of likely connected strings, then MC can calculate if indeed (statistically) the connection is given. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] re comments on Life and Death
On 04.09.2015 07:25, David Fotland wrote: group strength and connection information For this to work, group strength and connection status must be a) assessed meaningfully and b) applied meaningfully within a broader conceptual framework. What were your definitions for group strength and connection status, for what purposes did you use them and how did you apply them? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Number of Go positions computed modulo 2^64; source code available
How much computation time do you expect to reveil the complete exact 19x19 number? Or is more research necessary before I may ask this? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Kassandra talking
n 22.12.2014 09:46, Ingo Althöfer wrote: In total: Changing the random move generator typically will change the playing behaviour. However, it can not be well predicted if this change will be to the better or to the worse. Is this prediction theoretically impossible (why, under exactly which presuppositions) or is research in understanding it not advanced far enough? -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Move Evaluation in Go Using Deep Convolutional Neural Networks
On 20.12.2014 09:43, Stefan Kaitschick wrote: If MC has shown anything, it's that computationally, it's much easier to suggest a good move, than to evaluate the position. Such can only mean an improper understanding of positional judgement. Positional judgement depends on reading (or MC simulation of reading) but the reading has a much smaller computational complexity because localisation and quiescience apply. The major aspects of positional judgement are territory and influence. Evaluating influence is much easier than evaluating territory if one uses a partial influence concept: influence stone difference. Its major difficulty is the knowledge of which stones are alive or not, however, MC simulations applied to outside stones should be able to assess such with reasonable certainty fairly quickly. Hence, the major work of positional judgement is assessment of territory. See my book Positional Judgement 1 - Territory for that. By designing (heuristically or using a low level expert system) MC for its methods, territorial positional judgement by MC should be much faster than ordinary MC because much fewer simulations should do. However, it is not as elegant as ordinary MC because some expert knowledge is necessary or must be approximated heuristically. Needless to say, keep the computational complexity of this expert knowledge low. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
Re: [Computer-go] Move Evaluation in Go Using Deep Convolutional Neural Networks
On 20.12.2014 11:21, Detlef Schmicker wrote: it is not easy to get training data sets for an evaluation function?! You seem be asking for abundant data sets, e.g., with triples Position, Territory, Influence. Indeed, only dozens are available in the literature and need a bit of extra work. Hundreds of available local joseki positions do not fit your purpose, e.g., because also the Stone Difference matters there. However, I suggest a different approach: 1) One strong player (strong enough to be accurate +-1 point of territory when using his known judgement methods) creates a few examples, e.g., by taking the existing examples for territory and adding the influence stone difference. It should be only one player so that the values are created consistently. (If several players are involved, they should discuss and agree on their application of known methods.) 2) Code is implemented and produces sample data sets. 3) The same player judges how far off the sample data are from his own judgement. Thereby, training does not require many thousands of data sets. Instead it requires much of a strong player's time to accurately judge dozens of data sets. In theory, the player could be replaced by program judgement, but I wish happy development of the then necessary additional theory and algorithms! ;) As you see, I suggest human/program collaboration to accelerate program playing strength. Maybe 9p programs can be created without strong players' help, but then we will not understand much in terms of go theory why the programs will excel. For getting much understanding of go theory from programs, human/program collaboration will be necessary anyway. -- robert jasiek ___ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go