AlphaZero: What can we expect next?

It was first described in “Mastering the game of Go with deep neural networks and tree search” in Nature Vol. 529, released 2016-01-28. After that strong engines appeared: Crazy Stone Deep Learning (First Edition), DeepZenGo and Leela 0.6.

The second paper “Mastering the game of Go without human knowledge” was release 2017-10-19 in Nature Vol. 550. The information in that paper is the base of LeelaZero.

I wonder if this guy is thinking what a nice handle he picked in 2011?

FM AlephZero

[Note that aleph zero is actually the smallest infinite cardinality, the size of the set of natural numbers]

Google Makes Its Special A.I. Chips Available to Others, NYT, By CADE METZ FEB. 12, 2018 

https://www.nytimes.com/2018/02/12/technology/google-artificial-intelligence-chips.html

Yeah, now part of the google compute cloud service, adding to as many cores and GPUs as you like! What google has done is to be commended, because they have not merely entered competition with the hardware manufacturers (principally Intel and Nvidia), but it has produced hardware that is significantly superior to anything that previously existed for the sorts of computing that is needed for AI (basically large matrix calculations). Nvidia would like to argue with that, and their Titan V has the advantage (for some) of being available for purchase, but it looks like google is currently the leader.

One of the Alpha Zero main developer is an IM. I guess he is holding the games back so he can use himself some of the novelties A0 has come with.

Demis Hassabis was a very strong junior - world class - but he did not get an IM title. His FIDE rating is 2249 and he has had no activity since 2001. It is clear academic challenges became more important to him, and it is only recently that the two have sort of merged!

I used to play on the same 4NCL team as Demis, and I remember him saying this. He scaled back his chess much earlier than 2001 (and at the last team dinner I remember him saying that he was playing a lot more go - I believe he won or became equal first in the British go championships in 1997 or 1998) and very quickly got into programming, cybernetics, neuroscience, artificial intelligence etc. He founded his own games developing company (Elixir Studios) and had a contract with Eidos. He sold off the company and its immaterial rights before the age of 30, and began focusing more specifically on neuroscience and AI.

Yes, those who understand a bit more about the learning algorithms they used will understand that the details are not chosen by the human designers: the neural network generates probabilities for all legal moves and as it trains, these probabilities become more useful from its experience. The paper from DeepMind had much of interest in it.

Some things I would like them to have included are the exact learning algorithms, and the exact neural network architecture. I hypothesise that these may have been omitted for commercial reasons, which would be unwelcome if it were true, especially as the commercial value will require radical redevelopment of anything they have done to different areas. Indeed I doubt that much other than the hardware, development environment and the general philosophy of design can be preserved when applications move away from games. But I bet they will achieve something exciting in a more important application are such as medicine.

Yes, but that is entirely specific to go (determined by the 19x19 board plus the standard technique of padding for convolutional features), and it was superceded with AlphaZero even in the case of go (in the latest work, there were no features except the raw position used in go).

I am not aware of any information on the structure of the hidden layers of the neural network for chess used by AlphaZero.

[Mind fart deleted]

"As most of you already know, Google's AI company Deepmind, has created a computer program that plays Chess so beautifully it's left the world's entire Chess community sitting back in awe."

I don't "know" this at all and neither does any thinking, informed chess player.

That's because the "beautiful play" is done with the opponent at a time, etc. disadvantage.

The chess that has been published is beautiful: few knowledgeable people disagree.

[Mind fart deleted]