5 ms·
When you read Go strategy resources, you see a lot of things divided into what best practices were before AlphaGo and what they are now. It's a whole big thing.
by tptacek 3y ago
When you read Go strategy resources, you see a lot of things divided into what best practices were before AlphaGo and what they are now. It's a whole big thing.
It is still the case, though, that AI dominates humans at Go; humans didn't get so creative about the game that they put AI back on its toes (though some did discover exploitable AI "strategy bugs").
- pa7ch 3y agoAgreed, but I still think humans should get a little more credit for winning against AI no matter how. Its a competitive game with very simple and clear rules. A hole in AI strategy is a hole, even if quickly patched! I am still so impressed that Lee Sedol beat Alpha Go 1 game out of 5 way back when AI made its breakout. I was sad he felt so sheepish afterward for losing. In hindsight, I think it was an amazing accomplishment even if today an AI could beat Shin Jin-Seo (#1 player) 100 out of 100 times!
- paulcole 3y agoThis is true in Scrabble as well. When I was playing seriously there were strong players who played a ton over a board and had deep intuition about what made plays good and what made plays bad. In the late 1990s/ early 2000s there started to be a lot more in the way of computer simulation and analysis and some very strong computer players. One (general) example was that older players liked the idea of making longer plays using more tiles to "win" a race to the S and blank tiles (the best tiles in the bag). Computer simulations generally show that turnover (as this is called) isn't optimal and you're better off holding strong combinations of letters rather than playing them off hoping to draw something better. Now younger players are better than ever because all of their training came with the help of computer analysis and simulation. Of course in Scrabble a huge part of it comes down to just memorizing the words in the dictionary.
- cdelsolar 3y agoAI doesn't dominate people in Scrabble though. The best humans are better than the best AI.
- samatman 3y agoI wouldn't have expected that. Is it just a relative lack of interest in building an AI which can dominate Scrabble? It's a partial-information game, but the search space can't be as big as Go, and an AI has an advantage over human players in that the entire valid string family can be encoded into a trie or some other efficient data structure, it's never going to forget a word or think it can play one that isn't valid. My intuition is that AI should be able to crush the best human players at this point in time, but I'm open to being corrected on that if there's some aspect of the game which I'm not modeling correctly.
- ultrasaurus 3y agoThat strikes me as odd too -- but it might be because searching a dictionary is such an obvious computer advantage that it's not interesting to optimize. There are only 10 articles on arxiv.org that mention Scrabble vs 100s on Chess https://arxiv.org/search/?query=scrabble&searchtype=all&source=header https://arxiv.org/search/?query=scrabble&searchtype=all&sour...
- paulcole 3y agoJust finding the highest scoring word won’t make you all that good. If you played the highest scoring play available to you each turn you wouldn’t be that strong of a player. Maybe around top 200-500 or so in the US I would guess? And it’d be a super exploitable strategy by a decent player. The reason is that you need to apply some rules, like when to trade vs. making a play, balancing consonants and vowels for future plays, what parts of the board are too dangerous to make certain plays in, etc. It’s because of the distribution of unplayed tiles, the high-scoring spots on the board, and the 50-point bonus for using all of your tiles. Because of that, generally, you’ll do better by building towards a 50-point bonus play every 3 or 4 turns than by maxing your score on each turn. I’d be curious about letting a human player play with the assistance of the best bot available and seeing how much better that would make them. I guess part of the issue though is that in a 13 play game maybe 3 plays are meaningfully difficult. So it’d take awhile to see if the human is improving on the bot or not.
- 3y ago
- SonOfLilit 3y agoYou also see a similar division to 19th century and 20th century when a player called Go Seigen changed the way the game is player even more, I feel, than AI did (but don't take my word, at 7kyu I'm far from qualified to understand how professionals play)
- Alex3917 3y ago> When you read Go strategy resources, you see a lot of things divided into what best practices were before AlphaGo and what they are now. It's a whole big thing. Yes and no. The biggest takeaway from AI is that learning all the joseki doesn't actually matter that much, which has freed up players (except for the pros) to spend more of their time focusing on the more fun and interesting parts of the game. There are a lot of videos showing what josekis and strategies the AI recommends, but as a human you're likely not going to be any better off following them. This is for the same reason why AI analysis of fights is largely useless. That is, the reason why you lost the big fight (and the game) isn't that you didn't find that one obscure 9P move that could have saved you, but rather than you let yourself get cut 50 moves earlier. But the AI will never show you the move where you got cut as being the reason why you lost the game, it will only show you the one random move that you'd never in a million years actually be able to find. This video from Shygost sums up the most important strategy stuff that you actually need to know in order to get strong: https://www.youtube.com/watch?v=ig8cWuDSHTg https://www.youtube.com/watch?v=ig8cWuDSHTg
- nicklecompte 3y agoThe "strategy bugs" are a symptom of a more general shortcoming and why 2024 AI is still basically dumber than a mouse. Keel in mind that if you had a variation of Go where there was a "hole" in the middle of the board, both Lee Sedol and a competent amateur would be able to play competent "Doughnut Go" without any prior experience. But AlphaGo and its successors would certainly make a ton of dumb unforced errors unless it practiced at least a few hundred games. (I am basing this observation on similar experiments with a similar Breakout AI, not sure if these experiments have been done with Go.) Mammals, including humans, have advanced brains because we evolved to solve weird and unexpected problems with moderate reliability, not to optimize well-known benchmarks with high reliability. (This is also why plants are green instead of black.) By contrast, AlphaGo is a machine designed to solve a highly specific problem. The whole point of machines is that they dominate humans at specific tasks, otherwise we would just use a human. But we don't describe bulldozers as "superhuman" unless we're being intentionally obscure; the same should apply to AI. Otherwise we risk assuming the AI is capable of things it probably can't do without retraining.
- sandspar 3y ago"General-purpose" computer seems like a misnomer.