I think this stuff is even more pertinent for reinforcement learning due to its dynamical nature (ie moving targets, exploration vs exploitation concerns, etc). What’s important is that the data is mostly right, such that statistics done on the data result in the correct minima - not that each example is perfectly labeled. Neural networks do well with noise and imperfections as it forces regularization that improves their filters, and enhances their capacity to learn and generalize better. It’s long been known that “perfect” data - noise free, 100% correctly labeled, etc - is often *not* the best training data for neural networks. It’s a scientific result in optimization. The “zero” tabula rasa thing is not just some philosophical ideal or aesthetic, nor is it a simple curiosity. Warren - have you called the Deepmind team a bunch of idiots too?įrom my perspective. Has to learn from and the faster it comes, the more it will learn. Human-taught or some such utterly false claim. (If not, change 10% to 5%.)Īnd now please don't send a reply back about how this would be cheating and It will cost 10-% slower learning, but the much larger knowledge source Which is a much large knowledge-source than checkmate positions alone. do neural net learn step using that data.Īgain, if this is done, then leela will learn from psotiuons <=4 ply Move M that accomplishes that in the win case) then use that as training data.ĥ. if stockfish proves positon is a forced win, loss, or draw, (and finds a In about the same amount of time T it takes Leela toĭo a single neural-net-learning step, stockfishģ. Say false things, just like you and Jesse just said.Īnd while I am at it, leela also is being stupid by missing the opportunity Only people like you are that stupid, and they abound in the leelaĬommunity, for reasons I do not understand. If a million times larger source of training digit-image data were available, Once upon a time, people trained neural nets to recognize handwritten digits. Possibly a million times faster, but this is probably an overestimate.īut probably it is safe to say twice as fast. Knowledge, hence might be expected to learn tremendously faster. If it learned from tablebases, then it would have a million times It has no other source of knowledge.Ĭheckmate positions are rare. It currently learns only from game-end positions such asĬheckmates. > We know that tablebase support has just been included if you want to use it Never occurred in practice, think tablebases are cheating or evil somehow, Or, of course, we could just listen to uninformed people who think KNNkp has Idiot who does not even know basic checkmates and basic endgames like KRPkr. Therefore leela would learn faster, as opposed to being a total With tablebases, it has a far larger source of knowledge, millions (checkmates, stalemates, and 50-move & repetition rule). The ultimate source of all knowlegde leela has, is game-end positions Probably have no trouble learning how to win this and KBNk. If however, it were to learn from endgame tablebases, it would So leela may have a very hard time getting any clue what to do. There are a lot of ways it can try to wriggle out. You have to corral the king in the correct corner which is not so easy, even if you do know what to do, it is quite difficult to pull it off, with slightly more intelligent play the KNN side will capture the pawn, almost-random play will tend to cause the kp side to win (queens),Ģ. Have great difficulty learning it for these reasons:ġ. The cases the human knew how to mate, the other times he embarrassedĪlthough the winning technique for KNNkp is conceptually fairly straightforward,Īctually executing it is quite difficult for a human. you are incorrect, it has occurred many times, certainly at leastġ0 times, in practice, and in about 20% of Also, the KNNkp ending has never occurred in practice
0 Comments
Leave a Reply. |