< Earlier Kibitzing · PAGE 3 OF 3 ·
Later Kibitzing> 
Dec0917
  HeMateMe: This program "taught itself chess" then beat the present computer world champion? Perhaps this is the next step, software that can unlock medical and engineering mysteries. Chess is just an afterthought. It destroys GO champions too and masters every other game it learns. 

Dec0917   geniokov: Sooner or later,once this AI be converted into a small chips and be implanted into human body then all of us will be UPGRADED into new dimension in life. Hahaha! 

Dec0917   geniokov: The "End of Chess"is near once AI unlock all problems. 

Dec0917   ndg2: I wonder, since a0 taught itself chess, would there be types of positions, where it wouldn't fair so well, because they come from a different type of playing style? In other words, should it train is neural network,by playing against other engines, too, to make its own play more versatile? Maybe they thought about it, but felt it's already strong enough. And with alpha zero, it probably was exactly the point, that it's selftaught without outside influence. 

Dec0917
  alexmagnus: The developer of Stockfish himself said, if Stockfish played at its best (that is, in the best possible configuration), there would be <more draws>. He doesn't say anything about wins  apparently, even to him it is evident that AlphaZero is stronger. 

Dec0917
  Ron: <theagenbiteofinwit: I'm more interested in AlphaZero's opening choices, which we can learn from, rather than some of the weird choices it makes in the middle game?> A recent article in chessbase https://en.chessbase.com/post/thef... says AlphaZero is favoring the English Opening, and the Queen's Gambit. 

Dec0917
  Ron: From what I understand, AlphaZero has a learning algorithm.
That means that AlphaZero's level of playalbeit very highwill not remain stagnant; its ELO will increase the more it plays. 

Dec0917
  zanzibar: <Ron> f(x) = 1  exp(x)  x > 0 is a monotonically increasing function too. But it's also bounded. 

Dec0917
  WorstPlayerEver: Hmm... an interesting question: is there is a random factor involved? If there would be A0 and B0, exact copies running the same hardware, would they follow exactly the same path of development? And what if they play a match then? 

Dec0917
  WorstPlayerEver: PS: as a human being I'd rather like to know fer sure ;) 

Dec0917
  zanzibar: <WPE: If there would be A0 and B0, exact copies running the same hardware, would they follow exactly the same path of development?> No, as the method involves a Monte Carlo process  which involves random number generation. Presumably they generate a new seed for each trial. 

Dec0917
  Ron: I'm fascinated by AlphaZero, but I'm a bit concerned that the creators have only released the wins by AlphaZero. Releasing all the gamesall the drawn gameswould undoubtedly contribute to the growth of knowledge, including that of AlphaZero. That seems like the scientific way to me. 

Dec0917   waustad: Have they told it the axioms of mathematics and let it run from there? Maybe something with fairly simple rules like group theory would be a start. I wonder what its proofs would look like. 

Dec1017   MasterMito: there are hardware differences maybe 

Dec1017
  alexmagnus: <Have they told it the axioms of mathematics and let it run from there? Maybe something with fairly simple rules like group theory would be a start> The problem with using neural networks for mathematical discovery is, you need to formulate a goal, and a measure of how close to the goal the network got, as this is what eventually defines the network's learning and success. In chess, such a goal is easy  winning. What is a goal in an openend mathematical discovery? 

Dec1017   Chessinfinite: Astonishing ..
It 'learnt' or 'self taught' hundreds of years worth of human knowledge in a few hours.. this is scary if not the apocalypse just yet. 

Dec1017
  john barleycorn: <alexmagnus: ...
The problem with using neural networks for mathematical discovery is, you need to formulate a goal, ...> How about calculating the next 100 Mersenne primes? As warm up exercise. 

Dec1017
  HeMateMe: <How about calculating the next 100 Mersenne primes? As warm up exercise.> How about a math proof that convinces Merkel that Germany was already at full employment and fully populated and should not take in one million illegal refugees? 

Dec1017
  WorstPlayerEver: Well, well, it took Stockfish 4 days, but there's a new development version since the release of the stunning A0 performances. Which is interesting, because they released 11 DVs from 36 december 2017. While the last DV before these recent DVs is from 18 nov 2017... http://abrok.eu/stockfish/
In other words: Stockfish somehow *sensed something...* 

Dec1017
  john barleycorn: <HeMateMe: ...
How about a math proof that convinces Merkel ....> After a proof that HRC lost the election ... first things first. MAGA. 

Dec1017
  zanzibar: <<alexmagnus> What is a goal in an openend mathematical discovery?> Truth.
Godel built one of these neural nets a long time ago for such a purpose, but it got stuck in an infinite loop, or somesuch. . 

Dec1017   schweigzwang: <What is a goal in an openend mathematical discovery?> I reckon this is a fair question to ask humans, but not to ask AlphaZero. Maybe we should not be so quick to call AlphaZero "AI". 

Dec1317   dannygjk: So many uninformed comments. 

Dec1317   dannygjk: One of the chessgames.com tips says castle early and often. Well If you check Fischer's games he often castled later than many GM's. 

Dec1317
  alexmagnus: A0 doesn't castle late though.
A0's castlings:
6, 12, none*, 9, 10, 7, 8, 7, 6, 6
*moved king on move 16 



< Earlier Kibitzing · PAGE 3 OF 3 ·
Later Kibitzing> 