In December of this year (2018), Google released 1000 games of AlphaZero versus Stockfish – both chess programs. Until now Stockfish was the reigning king of chess programs, but AlphaZero is the new kid on the block it soundly beat Stockfish. Yes, that should scare all of us. It should scare all of us because it is the way AlphaZero works.
Deep Blue versus Gary Kaparov was the beginning of machine domination of the game. Today, the highest rated human player (Carlsen) is 2800+ and the highest rated program (Stockfish) is about 3400 to 3600! Stockfish would beat Carlsen every day of the week and twice on Sundays. It would not even be a competition. Such is the power of today’s chess engines.
Humans have played chess for over 1,000 years and today’s chess engines are finely tuned computing machines that have been coded with all the knowledge humans have gained about chess through these years. These engines have broad and deep opening databases, end-game strategies, they are fine-tuned by grandmasters, they understand piece valuation, and so on. For example, an engine knows that a rook is 5 points worth and a bishop or a knight is 3 points worth. An engine knows that certain opening favors white statistically and others not.
But then DeepMind – a company founded in London 5 years ago and now acquired by Google – went even further. They wrote the first artificial intelligence based program to play chess. It is called AlphaZero.
Unlike chess engines of today (like Stockfish), AlphaZero understands nothing about game. Let me say it again – it understands NOTHING. It only knows the rules of the game and conditions for winning (checkmate). It has no opening database, no end-game strategies, no grandmaster wisdom, no historical knowledge of which openings are better for white statistically, and so on. NADA.
So how does it get better?
It plays against itself to learn the game, using the rules of the game. In the beginning it makes random moves, makes horrible mistakes– just like a novice. But then soon it learns – just like a human and improves its game. Eventually, after playing a few million games, it becomes really really good. But you see, it takes very little time to play these games. Before playing StockFish, AlphaZero played against itself for only four hours! In four hours, it acquired all there is to learn to play the game. And it has soundly beaten Stockfish! It plays like no other human and it plays like no other engine. It makes moves we have not seen before and yet it wins the games against the most powerful engines of today.
The AlphaZero versus StockFish games are being analyzed on YouTube. In fours hours of play AlphaZero beat 1000 years of chess knowhow humans had acquired. And that should give all of us a pause.
But here is the real scary part. AlphaZero taught itself to play other games too! It has mastered GO – a Chinese game that is far more complex than chess, and most recently it mastered a Japanese game Shogi (another board game). AlphaZero is the world champion in GO, CHESS, and SHOGI. It has beaten world champions in all these three games.
AlphaZero uses deep neural networks and reinforcement learning AI techniques to learn – in fact teach itself – how to play games. I find that astonishing that all the human knowhow of chess and all the research in building great chess engines is soundly beaten by AlphaZero after four hours of training.
I am sure the field of AI has many such surprises for us.