i dived into your post and thought mr rattigan was a phiilosopher for a second.
then remembered. how about turkish ?can you speak turkish?
God knows what he is. I think a mixture between computer security specialist and joker.
My wife is waiting for me and I'll have to go in a minute. I've been enjoying talking to you. Regarding the people you were arguing with, nobody can help where they come from and Americans sometimes have less than usual tolerance for different cultures. They can't help it so don't take it to heart! Goodnight.
tamam
2 things.
first, Go is much more complex than chess. it took 20 years after deep blue to get the same level for AlphaGo.
second, a massive number of permutations doesnt necessarily mean that something cant be solved. checkers had 10 ^20 and was still solved. of course, chess is much much more complex, but the big number alone should dissuade us.
Your premise assumes that the efforts put into beating the world champs for Chess and Go were the same. This is not the case. Solving Chess was much better PR for IBM than solving Go would have been, so a lot more resources were brought to bear.
In terms of actually solving, IIRC Go has more positions...but evaluating Go positions should take less CPU horsepower than evaluating Chess positions.
evaluating go positions takes WAY more CPU than chess positions.
i recommend you look at the smithsonian article or the one by scientific american.
in terms of raw computing power,
evaluating go positions takes WAY more CPU than chess positions.
i recommend you look at the smithsonian article or the one by scientific american.
in terms of raw computing power,
I'll read the articles if you link them, but I did not find any in-depth articles about solving Go or Chess from either publication in my 5 minutes of digging that support your statement, so...I'll stick with the articles I have already read in the past on the subject.
My wife is waiting for me and I'll have to go in a minute. I've been enjoying talking to you. Regarding the people you were arguing with, nobody can help where they come from and Americans sometimes have less than usual tolerance for different cultures. They can't help it so don't take it to heart! Goodnight.
Americans don't really need you to be a spokesman for them, and if your premise were true, it begs the question of where did this "American" bias originate? Are there any other cultures known for running roughshod over every other culture in their way and appropriating them (and stealing their historic treasures for good measure)?
The only person whose biases you can speak for are your own, and there's plenty of fodder there to keep you busy for a good long time.
@7328
"start analyzing on Stockfish on your PC from the opening position,
and post back here when you get to 100 ply"
++ That takes 15,000 years on a desktop, or 5 years on 3 cloud engines of 10^9 nodes/s.
@7341
"you don't necessarily reach the EGTBs"
++ It is inevitable to reach the 7-men endgame table base at sufficient depth.
In ICCF WC Finals the average is 42 moves i.e. 84 ply.
evaluating go positions takes WAY more CPU than chess positions.
i recommend you look at the smithsonian article or the one by scientific american.
in terms of raw computing power,
I'll read the articles if you link them, but I did not find any in-depth articles about solving Go or Chess from either publication in my 5 minutes of digging that support your statement, so...I'll stick with the articles I have already read in the past on the subject.
https://www.scientificamerican.com/article/how-the-computer-beat-the-go-master/
https://www.nature.com/articles/nature16961
https://www.businessinsider.com/why-google-ai-game-go-is-harder-than-chess-2016-3
Those articles reinforce my point. A single Chess position should take more CPU power to evaluate than a single Go position. Chess engines still use deep searches and brute force calculation alongside machine learning, AlphaGo uses purely machine learning (though they did "cheat" and use human play to seed the process, unlike AlphaZero's machine learning...which ultimately will make AlphaGo's learned valuations imperfect in the end), with heuristics to decide on win probabilities, and the heuristics used are baked-in every time AlphaGo plays and learns...i.e. almost all the processing power would be already front-loaded and done before a new position is evaluated. The only evaluation that takes place for AlphaGo in a new position is "what does this position look like related to already evaluated positions, and what does the value network say is the best win probablility?"
The whole point of machine learning is that the previous AI work is subsumed into the valuation so that the next position *doesn't* require brute force calculation. All that AlphaGo calculates is "what worked best the last time a position like this came up?". It's like training a dog to do tricks, but the AI can remember a gazillion steps for its tricks and performs those steps perfectly every single time.
Machine learning for Chess works to a point, but as Stockfish has proven out, a combination of brute force calculation and machine learning probability valuations is stronger that machine learning alone. Which is inherently obvious if you ponder it for a minute or two.
If Chess has 400 possibilities after 2 moves, and Go has 130,000 possibilities after 2 moves, then if each single Go position took more CPU power than each single Chess position, AlphaGo would be running more than 325 times slower than AlphaZero on a given position using the same DeepMind hardware. I'm pretty sure that is not the case...but feel free to prove me wrong on that.
Ergo, the CPU usage for each individual Chess position > the CPU usage for each individual Go position.
Those articles reinforce my point. A single Chess position should take more CPU power to evaluate than a single Go position. Chess engines still use deep searches and brute force calculation alongside machine learning, AlphaGo uses purely machine learning (though they did "cheat" and use human play to seed the process, unlike AlphaZero's machine learning...which ultimately will make AlphaGo's learned valuations imperfect in the end), with heuristics to decide on win probabilities, and the heuristics used are baked-in every time AlphaGo plays and learns...i.e. almost all the processing power would be already front-loaded and done before a new position is evaluated. The only evaluation that takes place for AlphaGo in a new position is "what does this position look like related to already evaluated positions, and what does the value network say is the best win probablility?"
The whole point of machine learning is that the previous AI work is subsumed into the valuation so that the next position *doesn't* require brute force calculation. All that AlphaGo calculates is "what worked best the last time a position like this came up?". It's like training a dog to do tricks, but the AI can remember a gazillion steps for its tricks and performs those steps perfectly every single time.
Machine learning for Chess works to a point, but as Stockfish has proven out, a combination of brute force calculation and machine learning probability valuations is stronger that machine learning alone. Which is inherently obvious if you ponder it for a minute or two.
If Chess has 400 possibilities after 2 moves, and Go has 130,000 possibilities after 2 moves, then if each single Go position took more CPU power than each single Chess position, AlphaGo would be running more than 325 times slower than AlphaZero on a given position using the same DeepMind hardware. I'm pretty sure that is not the case...but feel free to prove me wrong on that.
Ergo, the CPU usage for each individual Chess position > the CPU usage for each individual Go position.
bro i think ur just straight misreading the articles at this point.
you.... do realize that the equivalent heuristic is that alphago evaluates a position 300 times weaker?
bro i think ur just straight misreading the articles at this point.
you.... do realize that the equivalent heuristic is that alphago evaluates a position 300 times weaker?
"Bro" can you even explain what you just said using your own words? Define "a position 300 times weaker". What does that mean to you? Or are you going to keep regurgitating? Do you even know how to program, or are you just reading these articles with no understanding? If someone told you to write a Chess or Go engine from scratch, how would you attack the problems?
@btickler you are making the false assumption that those engines are evaluating positions at the same strength.
im going to repeat myself here
"If Chess has 400 possibilities after 2 moves, and Go has 130,000 possibilities after 2 moves, then if each single Go position took more CPU power than each single Chess position, AlphaGo would be running more than 325 times slower than AlphaZero on a given position using the same DeepMind hardware. I'm pretty sure that is not the case...but feel free to prove me wrong on that."
this makes the assumption that the evaluations are of the same strength. they arent.
it is stated very explicitly "The game of Go has long been viewed as the most challenging of classic games for artificial intelligence owing to its enormous search space and the difficulty of evaluating board positions and moves."
"difficulty of evaluating board positions and moves"
compared to human - wise, alphago is the same level as deep blue.
deep blue had 11 Gflops. Alphago uses at least 84000. (1200 CPU*70 gflops each, assuming the cpu's are like those found in a regular computer.).
takes more than a thousand times the amount of power, let alone with better AI, for a Go program to perform as well as a chess program
@btickler you are making the false assumption that those engines are evaluating positions at the same strength.
im going to repeat myself here
"If Chess has 400 possibilities after 2 moves, and Go has 130,000 possibilities after 2 moves, then if each single Go position took more CPU power than each single Chess position, AlphaGo would be running more than 325 times slower than AlphaZero on a given position using the same DeepMind hardware. I'm pretty sure that is not the case...but feel free to prove me wrong on that."
this makes the assumption that the evaluations are of the same strength. they arent.
it is stated very explicitly "The game of Go has long been viewed as the most challenging of classic games for artificial intelligence owing to its enormous search space and the difficulty of evaluating board positions and moves."
"difficulty of evaluating board positions and moves"
compared to human - wise, alphago is the same level as deep blue.
deep blue had 11 Gflops. Alphago uses at least 84000. (1200 CPU*70 gflops each, assuming the cpu's are like those found in a regular computer.).
takes more than a thousand times the amount of power, let alone with better AI, for a Go program to perform as well as a chess program
No Sherlock, we're comparing AlphaGo to AlphaZero here, same hardware and two root branches of the same AI software. Why would you even bring up Deep Blue?
There's no assumption of "strength" required. I stated that Chess positions should take more CPU to evaluate than Go positions, one for one. Period. End stop. You have supported my position with every post you have made. The fact that you can't grok this is pretty funny.
" I stated that Chess positions should take more CPU to evaluate than Go positions, one for one." - which is objectively incorrect.
" I stated that Chess positions should take more CPU to evaluate than Go positions, one for one." - which is objectively incorrect.
Well, I've made a logical case. I don't see you putting anything forth other than blather and quotes you don't understand yourself but that have given you some vague notion that you must be right.
@7376
"Chess positions should take more CPU to evaluate than Go positions, one for one."
++ Solving Chess does not depend on some evaluation, but on the 7-men endgame table base.
Stockfish is designed to play, i.e. find one good move in some time limit e.g. 3 min / move.
To do that, it depends on some evaluation as it cannot calculate all the way in that time limit.
Stockfish can be used to analyse, using more time and taking e.g. 2 moves instead of 1 move into account. Then it will in part depend on evaluation, as some lines will not reach the 7-men endgame table base.
Stockfish can be used to weakly solve Chess, using much more time:
5 years on 3 cloud engines of 10^9 nodes/s, or 15000 years on a desktop,
calculating all the way to the 7-men endgame table base.
were you speaking turkish those days
I stayed in Adana in 1973 with a university friend who went on to become a professor of engineering, I think, maybe at Gaziantep, if I spelled it right. His sister started teaching me Turkish. It got me into trouble because a whole platoon of soldiers I was swimming with at Alanya started laughing at me and it turned out I was speaking Turkish with a woman's accent!