Life In 19x19 http://lifein19x19.com/ |
|
DH records http://lifein19x19.com/viewtopic.php?f=48&t=13520 |
Page 6 of 6 |
Author: | dhu163 [ Mon Apr 05, 2021 7:56 am ] |
Post subject: | Re: DH records |
won my last two games in the casual league against Javier A-Savolainen and Amir Fragman by resign after 40 point margins again. Came 2nd on 6/7. Valerii got 7/7! Well, I have been lucky enough so far to avoid coronavirus. Also lucky enough to start a new job project recently. Ever since AI, I felt improving at Go seemed more pointless and I have that feeling again. My training pre 5d (2013 Nov to 2017 ish): Lots of KGS games every day, several Go clubs weekly. I started with Nick Sibicky, later Dwyrin and then many hours of Chinese pro videos (tywq, weiqitv) every week (or even day), lots of NHK and KBS cup too without understanding the language. bought no books but did skim 1 or 2 I think, I'm not sure I thought of it as training, but more an addiction. I think I liked to record my games and theorise (nonsense) about what I was doing wrong. If I had a theory question, I read Sensei's library. I think I was relatively good at endgame and reading on the board with patience to calculate more, but weak at instinct and didn't have enough tsumego experience. I think I had some preliminary ideas to CGT without having heard of it, but was particularly curious about things that sound like they make sense but we can't prove like "play on wider direction etc." . Perhaps it was jarring to my experience of mathematical proof. 2017-2019 stalling at 5d: pretty much as above + AI reviews + watch lots of AI games, but less and less of everything. KGS, Pandanet games tending to zero, leaving only games at (very few) tournaments. Helped organise ICGT 2018, though looking back, I'm not sure I managed to contribute much. It was completely outside my normal experience and hence pretty interesting though. Passed (but terribly) in my degree. Some new (Taiwanese?) channels appeared on youtube in Chinese (e.g. GoGo) that I followed. Tried to program a small neural net to play tictactoe and failed. Tried to analyse neurons in leelazero and got too many bugs. Got an IT job. confidence boosting to reach this Go level. But unsure why the stall occurred when I was training like I used to. (perhaps because I was ??) 2019 Dec - Now: Nothing much in terms of training other than very heavy AI reviews after each game (like the whole next day), but somehow reached 6(-7d?) in strength though not officially. And keeping a scarily high winrate on PETC. Dangerously ego boosting. I have noticed the potential pun on the name of this thread. Stopped following youtube channels almost completely now. I still feel a bit weak at tsumego, but have included it in my routine more recently. My endgame has gone down the drain compared to my new skills in large scale fighting, especially in punishing my opponent's mistakes. But it remains hard to verbalise any of this instinct. I wrote a couple of "papers", available on LGC, but I feel they are pretty poor quality compared to what I could do, lacking in research, and containing flaws in thinking. -- I wasn't so sure of my direction. I dreamed of pro a little at one point, but I think it doesn't make sense now. But I think I would like to do some writing about Go. I'm not sure mentioning my goals here is so helpful, but it worked well enough last time (2017 British Champion ). TBH, I wrote around 40 pages of explaining maths theory of Go (+some new research) + 60 pages of mess around 1.5 years ago. But I haven't really got round to organising it into presentable form, nor clarifying the mess, nor verbalising new ideas. I have no excuse for not making full use of quiet time over the past year. Still, I find as I read more papers that my latest thoughts are often on the right lines towards the results in those papers. Perhaps I can still make progress. Hopefully my maths ability hasn't completely gone down the drain or else I will be an embarrassment to my username. So my goals in Go for the next year(s) in priority order: 1. Presentable book style of what I have written so far + extensions. Not sure about publishing. I'm not sure there will be much audience for such topics anyway, but I still think it's worth organising current knowledge into a coherent course sequence of material. I had never heard of most of it before I started investigating. 2. Do more research into existing mathematical research in Go. I have used only free resources so far, so I may consider buying some classics, such as Mathematical Go. 3. Do some Go research of my own if I can. 4. British Champion once more |
Author: | dhu163 [ Mon Nov 29, 2021 1:31 pm ] | ||
Post subject: | Re: DH records | ||
Won British championship. The long time settings are quite comfortable. My total point loss was around 60 points. (20220204: self-inflicted cross-cut, lazy stretching, arm exercise.)
|
Author: | Uberdude [ Mon Dec 06, 2021 12:16 am ] |
Post subject: | Re: DH records |
No comment on 79? That gave Bruno a big chance. Or did you think you could survive the ko even if objectively bad because you expected to outfight him and rather that than simple trade? Because he played d6 the ko became heavier for him with e4 corner trashing as another bonus for you, About left side cover, that happened in similar position next day in Nongshim cup https://www.go4go.net/go/games/sgfview/100268 |
Author: | dhu163 [ Sat Dec 18, 2021 12:57 pm ] |
Post subject: | Re: DH records |
Balance between assumption inertia and time recalculating. I probably assumed W had two ko threats on the right, thought I could respond to both and be fine, and so didn't recalculate after W's bad move 74. I did notice the possibility of ending the ko quickly in the post-game review, though I didn't yet realise that ending the ko was much better (that requires more precise calculation of values). I'm guessing you have Lee Sedol's move 78 (divine move) in the back of your mind. In Sino-Korean that's pronounced chil pal. 79 would be chil gu. Should I point out a number of disturbing coincidences in my go life at the moment? (rhetorical) One of the many is the round numbers in my rating points: 5, 5, 2500, 20. Are these caused by subconscious calculations, environment, intentional pranks or something else entirely? (edit:20220101. omg, even this post has weird references that I didn't notice then. Reading has weird effects on the brain. I must have been thinking of disturbing kos (a James Davies translation of an Ing rules term that I've seen in Jasiek's papers), just as my wordpress suggests I was thinking about goths.) |
Author: | dhu163 [ Sat Jan 01, 2022 6:38 pm ] |
Post subject: | Re: DH records |
well, I have written some things I am pleased with on my wordpress, and haven't yet noticed glaring errors, so I don't feel like a complete fail of a mathmo. otherwise, just an update: However the paper on loss-making ko threats that I was preparing is at 40pages and I'm not yet organised enough to handle the mess, especially with work too. I realise my solution + algorithm is only partial (or else I would have already self-published it), but probably worth publishing anyway, but I haven't even gotten round to writing up proofs of the most key statements yet. I would also like to investigate the cases it can't handle a bit more, and apply Tavernier's original simpler model to attack approach + multi-stage kos, but perhaps that should be another paper. I have found one of Bill's papers that uses a simplified Neutral Threat Environment to give counts of approach kos, and the appearance of Fibonacci numbers is amazing, though I don't fully understand the calculations. Random thoughts on bridge Some superstition Basic thought experiments for children, Brian Cox style: |
Author: | dhu163 [ Fri Feb 04, 2022 3:48 am ] |
Post subject: | Re: DH records |
Knot theory thoughts. Today's vote for word of the year is sao3 rao3 (disturbing/harassing) If you harass a knot for 1 trillion years, it might eventually unravel when the stars align. The grating feeling is anger. If you cut it, you break through, but the patterns remain. ------ [/Hide][/Hide]Dare eat a lion.[/Hide][/Hide]Just go around the valley. But there be wolves and eddies.[/Hide]Ida no wonderful Ing twa black awe.[/Hide] [/Hide] __ Update on goals. 20220316 I suspect it will take years to complete all of these. 1. ko paper (20hrs?) 2. constructing go trees board (400hrs?) 3. go book/writing (logic, concept classifications) (logic: 400hrs for organising, making more concise, deciding if more mathematical or not, better proof structures)(concepts: 200hrs total + absorption time). 4. go programming (influence functions: target hunters, cycles, interactions, short circuits) (??? is this even solvable ???) 5. organise my L19 writings, sgfs, notes on NTE (fibonacci bill ko)? __ QM: Double slit. This works in space too, so it isn't the air. Wheeler reminds us we have "it from bit", we set up the experiment consistently, get the same data consistently, but ala Feynman, our model of what is happening to produce the results guides our guess as to the next hypothesis to test. It is hard to explain exactly why measuring an electron in one slit causes the interference pattern to disappear. Does the measuring apparatus have "premonition" due to the presence of the electron beam gun and apparatus. The simplest explanation is that the electron itself has been affected (ala Fermi), and Copenhagen says its wavefunction collapses to a point upon interacting with the measuring apparatus. However, I am reminded of my mistaken calculation of the speed of a bubble rising through water since the water is also moving to compensate for the bubble. It seems likely that something similar happens with the electromagnetic field (some kind of ether), and perhaps Fermi would guess that the apparatus also has an electromagnetic field that repels part of that of the electron upon interaction. Just as the plate (with the slits) blocks the electrons path. Note the apparatus must change state from the interaction for us to notice it, perhaps producing a number at the end. This interaction with a tiny electron might not always be consistent, however well designed the apparatus is. When we have such consistent results, we suspect that a simple explanation should be the cause, but perhaps we can be reminded of the difficulty of Fermat's last theorem. Even the wave nature of the electron itself is counterintuitive to us (with or without any slits). Perhaps the effect of the electron going through the slits creates more particles that travel with it that cause cancelling effects of sinc^2. Note that summing an evenly sampled points of sinc^2 (with right period) gives the same total whichever alignment you start at. Probably not just a curiosity either. Are electrons even identical across space and time? Just because they have the same name doesn't mean they have the same history and all properties. This becomes like hidden variables which we know can't explain everything without action at a distance. |
Author: | dhu163 [ Tue May 24, 2022 4:31 pm ] | ||
Post subject: | Re: DH records | ||
PETC round 9. I played in the final round. Won after massive kill. I said in postgame review other than an opening blunder I couldn't think of major mistakes I made. AI says I was totally wrong. I probably don't have a PETC game where I lost more points than this (perhaps around 500?). This happens when the life and death of a big group is in question. Luckily it was my opponent's group. But it could have been a close game many times if my opponent handled well. Perhaps I wasn't fully focused, but it shows just how weak we both are compared to AI. I think it might be a sign I was playing too much by thought (as Bruno said) and not calculation nor instinct.
|
Author: | dhu163 [ Sun Jun 05, 2022 5:27 pm ] | ||
Post subject: | Re: DH records | ||
With the training I have done (much more than I planned) the past few months, I think I understand go a lot better in terms of words rather than by rote. I can't say I'm much stronger, still probably 6.5 +/- 3d. Here is a teaser for how slowly I follow AI variations to try to understand them. weak point theory algorithm simple tools Komi size seems to be highly related to the size of smallest living group possible. It is like a tense semeai in the corners. The first player wins the size of the smallest group there as well as more blocking on the outside too as they can expect at least half of the potential around due to control allowing attack.
|
Author: | dhu163 [ Tue Aug 02, 2022 6:05 pm ] | ||
Post subject: | Re: DH records | ||
It is easier to convey Go understanding in a review than in a game. This isn't quite a review as much as mostly copying katago's variations, but anyway. nb when Mr Jasiek says creator/preventor of endgame, if B is the creator (of follow ups) that means that there is a big move that occurs if B spends 2 moves in a row that wouldn't exist otherwise. Invariably (are there exceptions) this is because the area was originally W territory if B didn't play to create. But with move, follow up, B can get deeper breaking of W territory. reading his Endgame 4 endgame5 neural networks
|
Author: | dhu163 [ Thu Aug 18, 2022 12:29 pm ] |
Post subject: | Re: DH records |
Model of "perfect" algorithm for a Go position. some thoughts went through game 48 of invincible. What is the classic go style of Shusaku's era? |
Author: | xela [ Thu Aug 18, 2022 4:07 pm ] |
Post subject: | Re: DH records |
Interesting. I have two thoughts on this. 1. The history of both computer go and computer chess shows that programming computers to think like humans -- or, more accurately, to think like humans think we think -- hasn't gone well. A successful algorithm is one that plays to the computer's strengths. In fact, I'd argue that even the most analytical modes of human thought are far more intuitive and unstructured than we want to believe (when you lay out a "sequence of logical steps", how did you find each step?), so attempts to capture "human thought" in software have missed the mark from the very beginning, but that's a whole other discussion. Then again, in this context I don't know whether you mean "algorithm" as something you might implement in software, or a thought process for a human player to run, or an object of theoretical study... 2. Quote: Find large groups This is by far the most complicated and difficult part of the algorithm! Play san-ren-sei: do you have one large group now, or three small/medium groups? At some point in the middlegame you'll find you've crossed a line, but where to draw that line? |
Author: | lightvector [ Fri Aug 19, 2022 6:42 am ] |
Post subject: | Re: DH records |
GnuGo is basically what you get if you try to manually and laboriously hand-code human ways of thinking about Go into a computer. Like, if you look up the code (it's open source) as well as documentation and discussion from back when it was developed, you can see it actually does try to "identify groups" by taking individual chains and attempting to glom them together when they're "connected", and then evaluate each groups eyespace and strength. It computes radial territory and influence heuristics to estimate control over areas, it has a tactical pattern and good shape and tesuji database much of which was hand-coded, it does local reading to try to determine capturability and connectivity of stones, etc. Reaching somewhere in SDK is certainly not bad for such an approach. But quickly you run into the fact that all these heuristics are impossible to tune well, hand-coding them is crude and as xela said you have to draw so many arbitrary lines in arbitrary places, the complexity quickly balloons as you get all sorts of combinatorics that don't fit into any crude primitive attempt to categorize groups and such (e.g. combinatorics where you have miai to connect between groups, but getting cut results in different costs or implications for territory/influence/future forcing moves, or where connection is contingent on ko, etc - are they the same group or not? It depends, often you can't simply categorize it like that). Which is why GnuGo has never gotten past where it is. Hand-coding expert heuristics simply doesn't scale for making a strong computer Go player. Although maybe trying to think more systematically or quantitatively could certainly help some human players, if your focus is on finding better algorithms or approaches for humans rather than for computers. |
Author: | dhu163 [ Fri Aug 19, 2022 1:35 pm ] |
Post subject: | Re: DH records |
Thanks for the clarifications. My writing has been a bit careless. However, I think that any good system needs to incorporate groups, eyespace, connection, value etc. And that such should be found within neural networks if someone figures out how to look. I have read some of GnuGo. My feeling is that GnuGo perhaps needs some endgame theory for sharper evaluation. And this sort of endgame-like theory maybe hasn't been developed yet. Sanrensei sort of counts as a large group with overlapping influence of each stone, but enough space that there is much to fight for. It needs to do some minimax of the value of the cutting points, development areas and how that changes with the surroundings. Perhaps it starts looking like Feynman diagrams ... . The ideal is to be able to factor into independent problems and use averages (perturbation theory etc.) when dependent. If the increasing temperatures of fights can be related to sums of divergent series (1+2+...=-1/12), then I would be amazed. Go is a difficult enough game that we expect any complete algorithm to be computationally expensive. But I have some faith that a physics like theory exists that can answer many general questions about principles, etc., and some even with high precision. |
Author: | xela [ Fri Aug 19, 2022 6:37 pm ] |
Post subject: | Re: DH records |
dhu163 wrote: I think that any good system needs to incorporate groups, eyespace, connection, value etc. And that such should be found within neural networks if someone figures out how to look. I agree with this much. But I think the system incorporates these things as emergent phenomena from the training process, not from a human listing them and putting them in. And I think they're already there in the strong neural nets, but we don't yet know how to look for them. Think again about how you'd teach another human to play go. After explaining the rules, you don't just chant slogans at them. You need to play a bunch of games with them (give them some training data to work with). Concepts like groups and eyespace won't make sense until they've had enough experience to more or less form the idea for themself. You can name things, and use words to nudge the learning process in the right direction, but you can't directly implant the concept into someone else's brain. I did have a go at looking for the concepts of vital point, ladder and temperature. If I could be a full-time go/AI researcher, I'd like to more of this. But real life (and other hobbies!) keep getting in the way... Don't let me put you off. I'm interested to see where you're going with this. |
Author: | dhu163 [ Sat Aug 20, 2022 5:47 am ] |
Post subject: | Re: DH records |
ok, thanks. I mostly only have informal theory for now. Why is keima good for attacking? The local key attack points tend to be the same for dragon (large weak group). This is because it dominates the local value equation, regardless of how much profit the attacker can obtain from the other side (when it stops dominating, it may be sacrificed). In particular, the exponent of 1/2, the number of moves the attacker must play to kill, dominates. So getting say 1.5x faster to kill is important. Keima stretches the furthest locally. There may be a cutting point, but the cutting stones are weaker than the attacking stones. For example, with the ogeima, at least one cutting stone tends to be stronger than one attacking stones, so it is risky, and often better when the attacker is too weak to attack and just wants to live. And the attacker depends on getting the next attacking move locally (also very valuable) if the dragon cuts, which also strengthens the weaker of the attacker's stones. There is the possibility that this leads to semeai if the attacker's original group is weak too, in which case the attacker may consider that the keima might be too much without preparation. In general, consider the balance between value locally. Where are the weak points, i.e. the sources of value and how strongly do they pull? Like gravity? Some are chained. How does chaining weaken their value. It delays by powers of 2 according to number of moves required, dividing board per moves. What is a fight? The value of some weak group A is being fought over. However, the opponent has a weakness and weak chain(s) at B,C,D nearby (where B is weaker than C than D). A could play defensively at the boundary of A/B to live with follow ups, or cut off B, trying to either get two moves to take B, or move into reducing the potential of C,D. The temperature has increased from size of A+AB, to the size of A+B+BC (the potential at the boundary of B,C). If the attacker continues trying to rescue B, then the fight moves towards weakening C, etc. Perhaps B was 2 moves from being captured and C was 4. However, when B saves itself, perhaps C goes down to 2 and this spirals onwards. How to deal with open space in the centre? I don't know. This is like lots of little fights over every intersection. Too difficult to understand without averaging. _______ 20220825 Just thought about xela's explanation of fpu. That does seem like a good general algorithm (zero knowledge) but it puts a lot of trust in the neural network. I think that splitting the fpu bound into a sum of local katago expected control values should improve error significantly, but it may be quite a bit slower for a first instinct. But it should prevent rashes which occur due to overdependence on the policy's quality. The net is more motivated to improve the top policy suggestion as much as possible and less for the others. Perhaps there is a reason they call it 'alpha'go. Probably the biggest difficulty is working out how to implement this idea. If the policy top move isn't good enough I think that it us difficult for these neural nets to find the granularity to zoom in at the differences between the other options. A combination of general algorithm with domain knowledge probably helps. As well as focus on evaluation of weak points in Go. I think xela has pointed out the key reason why bots are so bad at calculation and start flailing around in the dark when they make a mistake. It can look as if they are on tilt. But at least it makes itself obvious. Because they don't have as strong a mechanism for a backup plan. I trust that Deepmind are aware of this and have fixed it for more important products. Perhaps this is a part of what they meant when they said alphago wasn't even in beta. 20220924 Today my vote for my word of the year is 忍 ren3, to endure rather than overplaying or ignoring when your opponent has a strong attack that you may have missed. |
Author: | dhu163 [ Mon Nov 21, 2022 4:12 pm ] | ||
Post subject: | Re: DH records | ||
Day before PETC: A random other point. Playing on KGS again, my rank was quite stuck at 4d initially, quite mysteriously. But now I realise similar things have happened repeatedly in the past, whether in things I've said or even my 4th place at EGF Grand Prix 2020. Probably my code-breaking experience links 4 to D, and something is mocking me, maybe even myself? Perhaps me not updating my rank on L19 has had more effect than I realised? Weird that geophysics seems to also have started coming into the picture more lately. Probably I used AI too much, regretting the pointless energy cost, as well as inherent interest. I have stopped using it for about 2 months, and plan not to as long as possible. Only IGS reviews (new cool feature) and reviewing online KataGo games. Mood: disorganised. writing lots of relatively long pieces on different topics rapidly, but not really understanding where it is going. PETC: Looking at the score estimate mess, perhaps it means white represents just the tip of the iceberg. Hmm, maybe the w shape represents a pig toilet not just a train line? Centrifuge. It does look like something spinning, perhaps tornado or fusion reactor. No, I still feel as though it is part of a circle meeting a square. Whether spinning or not, there are layers in between them that are shed.
|
Author: | dhu163 [ Thu Jan 05, 2023 10:16 am ] | ||||
Post subject: | Re: DH records | ||||
probably much more going on than I can describe. Of the things I've noticed, perhaps some were nice Christmas presents. "Am I too late to register?" "No, but I'll be angry if you don't play in all your games. Daniel Hu, 5d, No club?" "OK." ***silence*** **So that means I get to act as a 5d rather than 7d, and mustn't use AI?** I can't clearly remember the order of moves. public: more private: creativity vs flexibility
|
Author: | dhu163 [ Thu Jan 05, 2023 10:21 am ] | ||
Post subject: | Re: DH records | ||
unable to attend the last 3 games. 20230318: A review of a game I lost badly against KataGo on KGS. It feels as though they let me win the game after that, though perhaps I did learn something. As usual I review the opening way too deeply and struggle to get to move 10. I may add more later. As usual my understanding is numerical wrt AI's suggestions for optimal play. Hmm, it says "could not upload attachment."
|
Page 6 of 6 | All times are UTC - 8 hours [ DST ] |
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |