Laboratoire de Physique Théorique
Toulouse - UMR 5152

Computers are more and more present in everyday life, and they carry out more and more tasks previously reserved for humans. In particular, recent developments in artificial intelligence have shown that many decision-making situations can be handled by computers in a way that is at least as effective as by human beings : metro driving, autonomous vehicles, etc. However, the processes used by computers are very different from human processes, and it is important to understand these differences to assess the risks and limitations of artificial intelligence. In 1950, the English mathematician Alan Turing proposed a famous test to highlight these differences : a human being in an isolated room must ask questions to two interlocutors, and by the answers must determine which one of them is a machine.

A particularly spectacular application of artificial intelligence corresponds to board games ; For twenty years, computers have managed to beat the best chess players, but for the game of go the level of professionals remained inaccessible to the best programs until recently. It is only in 2016 and 2017 that the AlphaGo program, using recent advances in artificial intelligence, has been able to beat go champions repeatedly.

In an article recently published in Europhysics Letters, Célestin Coquidé and Bertrand Georgeot of the Laboratory of Theoretical Physics (University of Toulouse -UPS / CNRS) and Olivier Giraud of the Laboratory of Theoretical Physics and Statistical Models (Université Paris-Saclay / CNRS) applied the network theory to this problem. From a database of thousands of games played by professionals, as well as games played by go simulation programs, they built a network describing the structure of relationships between successive moves on the game board. They showed that the structure of the networks coming from the human and computer parts was different. This difference can be quantified from statistical indicators, which allow from a moderate number of games (a few thousand) to differentiate between human players and computers (see figure). This represents the equivalent of a Turing test for go simulators, which in addition makes it possible to distinguish the different types of simulators between them.

From the theory of networks applied to the game of go, the researchers could define two statistical quantities (represented on the two axes of the figure) which can be calculated with a few thousands of parts, and distinguish computers from humans : the points corresponding to the human / computer differences are well separated from the points corresponding to the differences between groups of human parts and groups of parts played by computer. Three computer programs corresponding to three different types of simulators were tested : Gnugo, Fuego and the famous AlphaGo.