Reading time:

Google's AlphaZero Chess Computer beats Stockfish with a 3600 Elo Performance

The world has a new chess champion it's called AlphaZero and created by Google/Alphabet Group developers in the DeepMind research team.

The major breakthrough idea for the Machine Learning system was Deep Reinforcement Learning, the brainchild of Matthew Lai. Lai wrote a Master's thesis on the subject matter at Imperial College, my alma mater, in which he conceived of a neural network learning system for Chess he called Giraffe. The system had modest success and became a pathfinder for the record-breaking AlphaZero chess computer.

AlphaZero does not use any domain knowledge like openings or endgame tables according to the developers. It uses perfect knowledge of the rules of the game and learns it from scratch. It uses no human input and plays itself honing its neural network by a method called Monte-Carlo tree search (MCTS). This is a completely new approach from the Negamax Alph-Beta Pruning tree search

The alpha-beta pruning and other innovations such as bitboard had propelled chess computer programmes or engines, like Stockfish, Houdini, and Komodo into a superhuman league of their own with Elo ratings of 2300 being the norm.

Alphazero uses a "neural network landscape" to represent the board and its evaluations. It iteratively improves its approach to the game by a results-based reward system using MCTS. The performance of Alphazero is now in excess of 3600 Elo compared to Stockfish Chess, a former reigning World Computer Chess Champion.

Demis Hassabis, one of the DeepMind team behind AlphaZero, tweeted.

no opening book, no endgame database, no heuristics, no nothing! full paper coming soon, will have things like early games.

— Demis Hassabis (@demishassabis) December 6, 2017

So what does this all mean for chess in general? Well, it has vindicated the proponents of "Hard AI" and cast doubt on the use of human-based heuristics in favour of zero-based learning systems. However don't expect Magnus Carlsen to play a match against Alphazero anytime soon, nor expect ChessBase to develop a boxed version of Alphazero for download.

Daylen Yang, member of the Stockfish Chess Team responded.

re: the AlphaZero achievement. super impressive. but also, it's trained on (my estimate) $20M of hardware (5000 TPUs). Only large industry labs like Google or Facebook even have the resources to conduct this type of research.

— Daylen Yang (@daylenyang) December 6, 2017

Many humans have given up trying to compete with computers and now they just got stronger it is time to harness them as Kasparov discussed in his book Deep Thinking. Guardian journalist, John Naughton made the following observation in his editorial review of Kasparov's Deep Thinking book.

In the grand scheme of things, losing by one game in a six-game match might not seem much, but at the time it was seen as a major milestone in the long march towards “artificial” intelligence (AI). With the 20/20 vision of hindsight we can view it in a less apocalyptic light: the triumph of Deep Blue was really a victory of brute computing power, clever programming and the ruthless determination of a huge but struggling corporation to exploit the PR advantages of having one of its products do something that would impress the world’s media. But if you believe that AI has something to do with cognition, then Kasparov’s epochal defeat looks like a sideshow. That it retains its fascination owes more to the popular view of proficiency at chess as a proxy for superintelligence rather than as possession of a very specialised skill. We’ve known for centuries that machines are much better at some things than we are. That’s why Google has become a memory prosthesis for humanity and why we use power drills to anchor bookshelves to walls. So the fact that machines now play better chess than even the greatest grandmasters or that DeepMind’s AlphaGo defeated the world Go champion at his particular speciality is interesting – and might even be useful in other areas, such as pattern-matching. But it’s just an incremental step on the same path that Deep Blue trod: the IBM machine used brute-force search; AlphaGo combined even more powerful brute-force search with a couple of neural networks. It’s technically sweet, certainly, but of less than cosmic significance. Sign up for the Bookmarks email Read more Living, as we do, in a time when existential concern about “superintelligence” and robots taking away middle-class jobs, Kasparov has acquired a new significance as the highest-profile (and highest-status) human ever to have been defeated by a machine. (Interestingly, Deep Blue didn’t take away his job: he continued to hold the world chess championship until his defeat by Vladimir Kramnik in 2000. And he continued to win tournaments and maintain his world ranking until he retired in 2005.) So what makes his book fascinating is that he uses it to reflect on what it was like to have been defeated by a machine and on the more general implications of that experience. The Kasparov v Deep Blue match has been endlessly discussed by chess aficionados in books and articles, but Deep Thinking gives us the inside story of what happened. Even for readers with only a passing interest in chess, it’s an absorbing, page-turning thriller that weaves a personal account of intellectual combat with the wider picture of what it’s like to come up against a powerful corporation that is determined to do whatever it takes to crush opposition. So this isn’t just a tale of human versus machine – it’s also a story about one man versus The Man.

AlphaZero is the story about one neural network against many computers, humans, and history itself. AlphaZero: 1, Everyone else: 0.








United States