AlphaZero: What can we expect next?

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. While the word "Genius" is always reserved for such brilliance, in this case it seems to have been relegated to an inferior option for describing just how great the artificial thinker actually is.

While it might seem like a logical choice for GM's and other great players to implement into their own games some of AlphaZero's plans for chess board domination, it's possible that humans simply do not possess the brand of intellectual depth required to follow through with those ideas. I'm sure many will try resulting in hit and miss results.

Deepmind has released only 10 of the 1300 games.

- 100 games from the starting position: (28 wins, 72 draws, 0 losses.)

- Twelve, 100 game matches from predetermined positions: (290 wins, 886 draws, 24 losses.)

The question is, why? For what reason are they holding back the other games?

We've been told that they're waiting to produce a full report on the match before releasing the other game. While that's probably true, I believe they're working on something else right now.

My speculation is that when Deepmind gets around to releasing the other games, they will also release news that AlphaZero has produced a repertoire for the white pieces (and possibly the black) that never loses. I believe they're trying to solve chess. Why not? They already did it with the ancient Chinese game of Go.

I believe however that they wont attempt to do this through a single set of moves that never looses, but rather a choice of systems where AlphaZero will always have a response to any move by the black pieces which will always lead to a white victory.

Based upon AlphaZero's performance with the white pieces, I'm fairly confident that Deepmind has pondered this possibility. It's likely they're attempting to narrow down the options into superior openings and superior lines that will eliminate the draw.

Doing the same thing with the black pieces seems like a tougher task. But with the white pieces I think it's within reach.

At any rate, I'm really looking forward to the release of the other games. I love the Chess.com analysis and commentary, especially by Daniel. 

Training tools like the one they released for Go: https://alphagoteach.deepmind.com/

Interesting. I wasn't aware they offered that. Thx!

DeepMind didn't solved Go, in fact they don't know what the limit of AlphaGo Zero or AlphaZero is. The training of AlphaGo Zero was aborted after 40 days. It still became stronger, but they said they need the hardware for other tasks.

I never thought about it, but the win rate of AZ vs AZ games for white and black would be interesting.

In my opinion DeepMind gave the impulse for the chess community to experiment and develop new chess engines. And maybe a NN can be trained to solve chess problems? (This is a spontaneous idea, it's possible that the currently existing NN is strong enough and a special NN is nonsense.) What happens if a 7 piece end game database is used during training and in the playouts?

In Go AlphaGo influenced the style of professional players which experiment with new moves played by AlphaGo. Because DeepMind gave some statistics of the openings played, liked and disliked by AlphaZero, this will have an influence on chess players too?

Something seems fishy about all that indeed. They don't disclose the games, they don't disclose their engine for public use, their engine doesn't appear in the engine ranking, even tho it has potential to be in first. But most importantly, why did they take so long to make a chess engine if they already made a go engine? If deep mind is truly neural and if they only need the rules of the game and the final result as input (after self training), why don't they apply that to a variety of different games?

They already disclose a lot: the basic concept and NN structure, games, multiple statistics. I lasted two years to develop AlphaGo Fun (end of year 2015) to AlphaGo Zero (October 2017). If I remember correctly there were three papers, two events (matches against Lee Sedol and Ke Jie), there was a variant of AlphaGo Master on a public Go server (which won 60 of 60 games against professional players on Tygem), and the AlphaGo Teach tool was released. In total 213 games played by all AlphaGo versions from 2015 to 2017 are available.

What is missing are the latest games of AlphaZero. Go and Shogi games are completely missing, for chess at least 90 games are missing (or 1210 if you count the games started with fixed openings).

I hope DeepMind will release the 90 chess games. I don't see any reason why they are hold back. There must be a reason of course, maybe DeepMind simply does not care about interested chess players. This can be disliked and criticised, but there is nothing fishy about it. They released a lot of data before, and AlphaZero is clearly a derivation of previous works. That a completely self learning AlphaGo-like NN works is proofed.

Unfortunately, I believe DeepMind has already moved on from chess, and we probably won't see any more AlphaZero games in the future. (I still hope I'm wrong on this, though!)

The AZ v. SF match was just another rules-based platform for DeepMind to test their AI's self-learning abilities.

The AZ technology will likely be tested on other things now—with the eventual goal (according to DeepMind's CEO) of using the AI to discover new treatments and cures in medicine.

Here is a good pointer to where DeepMind is going next.

http://blogs.bmj.com/emj/2017/11/08/how-theme-park-space-invaders-and-go-have-paved-the-way-for-exponential-healthcare/

Since it is not their goal to make just a chess engine, any further work on AlphaZero needs to be something that has yet to be done by AlphaGo Zero (beating the strongest players at the more computationally difficult game, by self learning) - no point wasting more millions to do the same work. Most likely the interesting next steps would be a whole new genre like RTS games, which will be a natural continuation to vehicle driving.

Didn’t they want to create the strongest Starcraft player? As far as I know this is more difficult because in chess or Go you have a clear final state, while in Starcraft a single decision does not result in a win or a loss. Multiple things can happen at the same time, because it is RTS. And t is not possible to simulate the game to the end, or at least the result does not say much about a decision, and the simulation cannot be done fast.

The goal of Alphazero was self learning, not furthering chess. It likely won't play chess again. Aphlazero won a different type of game after winning against Stockfish. It may have overpowered Stockfish with 1000x as many processors since Stockfish was not designed to use that many, I heard. I still wonder why they did not release the other games. Are those 10 the strongest wins? And why not a few games where Stockfish has the opening book and the endgame tables?

What would be interesting to: see some games were AlphaZero plays against itself. Not training games, but games played with the final NN.

I wonder how often it wins?

Or give AlphaZero 4 processor cores, same as what Stockfish is optimized for. Google can't say they don't have the hardware for that!

Stockfish is best with limited hardware. AlphaZero makes use of powerful hardware extremely well (the hardware would be useless to Stockfish, as it does not do large matrix calculations). AlphaZero plays better chess, not better chess with limited hardware.

Am wondering if Alpha Zero or something similar could be used for investing in stocks?  If possible--some people could make a lot of noney.

The stock market has far, far less information in it than chess and a great deal of randomness.

I've zoomed in and out on time scales in the stock market, went backwards in time, etc, and asked how I would have bought or sold with the info at the time, and whether it would have paid out looking at what happened. You can look at old data and test yourself with it. My conclusion is that it is very random and random on many levels. If you can wait for 10 years for a stock drop to come back up, you can make 7% per year on average with index stocks, vs 3% with treasury notes, or you can try your luck with your own business. Beyond that, watch the trading fees, and hopefully don't get taken by people with insider knowledge, encouraging you to buy because they really want to sell.

Even science is more wide open with possibilities than chess is. Wrong experiments in a lab make explosions. Done right they are not as cheap as a chess simulation. Alphazero is just a step forward. Much work is ahead still.