10 December 2017

Giraffe and AlphaZero

Start with the sociology of chess, last seen two weeks ago in FIDE's Social Commissions 2017, and add artificial intelligence, as in the previous post, A New Style of Chess, about Google/DeepMind’s AlphaZero. What have you got? I didn't know, so I asked Google.

The first answer it gave me (in fact, the first three answers) was a paper by Nathan Ensmenger: Is chess the drosophila of artificial intelligence? A social history of an algorithm. (ncbi.nlm.nih.gov/pubmed; February 2012):-

Abstract: Since the mid 1960s, researchers in computer science have famously referred to chess as the 'drosophila' of artificial intelligence (AI). What they seem to mean by this is that chess, like the common fruit fly, is an accessible, familiar, and relatively simple experimental technology that nonetheless can be used productively to produce valid knowledge about other, more complex systems. But for historians of science and technology, the analogy between chess and drosophila assumes a larger significance.

As Robert Kohler has ably described, the decision to adopt drosophila as the organism of choice for genetics research had far-reaching implications for the development of 20th century biology. In a similar manner, the decision to focus on chess as the measure of both human and computer intelligence had important and unintended consequences for AI research.

This paper explores the emergence of chess as an experimental technology, its significance in the developing research practices of the AI community, and the unique ways in which the decision to focus on chess shaped the program of AI research in the decade of the 1970s. More broadly, it attempts to open up the virtual black box of computer software -- and of computer games in particular -- to the scrutiny of historical and sociological analysis.

A little further down the list of Google's results was An AI computer learned how to beat almost anyone at chess in 72 hours (qz.com; September 2015):-

Matthew Lai, a computer scientist at University College London, recently published his master’s thesis, which demonstrated a machine learning system -- called Giraffe, after this cartoon about evolution -- that can learn to play at the International Master level of chess in just 72 hours. According to MIT Technology Review, Lai’s machine is a deep neural network -- a computer system that’s inspired by the structure of the brain and attempts to learn and make decisions in a similar way. According to Lai’s paper, Giraffe performs "moderately better" than contemporary computer programs that analyze every possible move at once, as opposed to the few that might actually lead to success.

Matthew Lai and Giraffe. Where have we seen those names recently? They were in the paper that announced AlphaZero to the world: Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm (PDF). Matthew Lai was listed as one of the 13 authors of the paper and received a mention in the references:-

Matthew Lai. Giraffe: Using deep reinforcement learning to play chess. Master’s thesis, Imperial College London, 2015.

Giraffe was mentioned again in the section titled 'PriorWork on Computer Chess and Shogi':-

Giraffe evaluated positions by a neural network that included mobility maps and attack and defend maps describing the lowest valued attacker and defender of each square. It was trained by self-play using TD(leaf), also reaching a standard of play comparable to international masters.

Looks like we're on the right track:-

Chess + Sociology + AI => Matthew Lai + Giraffe => AlphaZero

Where to take the subject from here? That will have to wait for another post.

No comments: