![chess-3325010_960_720.jpg](https://ipfs.busy.org/ipfs/QmPZs27zVWxdQmefeTYppjhHiEPD38AsuvjwYyZrcF7K31)

In early 2017 AlphaZero hit the stage by learning go, chess and shogi.  Let me emphasize the word ***learned***.  That's right AlphaZero learned how to play go, chess and shogi using a general reinforcement algorithm (essentially maximizing reward for making a decision.)  It learned how to play these games and beat some of humanities best.  This all resurfaced with Garry Kasparov's recent article in *Science* magazine called - [Chess, a *Drosophila* of reasoning](http://science.sciencemag.org/content/362/6419/1087)

>Garry Kasparov is a Russian chess grandmaster, former world chess champion, writer, and political activist, whom many consider to be the greatest chess player of all time.

(Wikipedia)

So, it just used computing power to figure out all the moves, right?  Well, hold on.  Sure, typically these system are straight bruit force probablility calculations on hundreds or thousands of possible moves. The next move is the next best move. However, AlphaZero was trained using a heuristic algorithm called the Monte-Carlo Tree Search[^1] (MCTS).  And here is where is gets interesting.  Usually deep learning systems are trained by using supervised learning from human experts (data from human played games for instance), and then by reinforcement learning with self-play.

What they did this time was to setup the game play rules and allow AlphaZero to play itself over and over becoming its own teacher. The neural network predicts its own move and impoves the tree.  Doing so helps to make higher quality moves and stronger play with each iteration. AlhpaZero started from scratch - hence the "Zero" and ended up beating Stockfish 8, the best chess playing program out there (it was able to beat grandmaster Hikaru Nakamura in 2014) by besting it 100 games to 0 (all winds or draws.)  This was after four hours of learning by AlphaZero. 

I thought that some of the descriptions by well known human chess champions were interesting and extremely erie.  They talk about AlphaZero having it's own style.  Even to go so far as to say it appeared to toy with or tease one opponent. 

Matthew Sadler an English chess grandmaster had this to say:
> “Impressively, it manages to impose its style of play across a very wide range of positions and openings,” says Matthew, who also observes that it plays in a very deliberate style from its first move with a “very human sense of consistent purpose”.[^2]

https://youtu.be/pFtY7gNRVRI

Ultimately what does this all mean. So what? Who cares? A computer program can learn and play a game of chess.  Big whoop, right?  Well not so. Playing games gives developers and scientisits environments that simplify the scope of real-world problems and at the same time retain complexity that challenges both humans and learning systems. IBM Deepmind is still learning and analysing the results and descissions that AlphaZero made during the games with Stockfish. 

This is all heading towards a general purpose learning system that will find answers to many of the complex probelms facing scientific communities. IBM has extended the Alpha platform to AlphaFold.  AlphaFold uses its knowledge to build 3D models of proteins[^3].  AlphaFold is a huge advacement and would be used to help unfold protiens in deseases, caused by folded proteins, like Alzheimer's, Parkinson's and other disease of this nature. 

What do you think about AlphaZero's nonhuman learning approach? 

![allthebest-future.png](https://ipfs.busy.org/ipfs/QmTinZ5vnG64GyBxFr4iU5YMTw2RmvvWJQDkfucLtfBqeM)


More Resources:
https://venturebeat.com/2018/12/06/google-deepmind-alphazero-chess-shogi-go/
https://deepmind.com/blog/alphazero-shedding-new-light-grand-games-chess-shogi-and-go/
http://science.sciencemag.org/content/362/6419/1087
https://en.wikipedia.org/wiki/Garry_Kasparov


[^1]: https://en.wikipedia.org/wiki/Monte_Carlo_tree_search
[^2]: https://deepmind.com/blog/alphazero-shedding-new-light-grand-games-chess-shogi-and-go/
[^3]: https://deepmind.com/blog/alphafold/