In this article we discuss: How chess and AI mutually evolved and where this has taken us. The (limited) role of much of the AI as well as the skeptical chess community in this evolution. We also discuss the relevance of the concept of “concept inflation”
Half a century ago, Dr. Milan Vidmar, a renowned Slovenian inventor and a world-famous chess player, made a somewhat humble observation to the effect that “concept inflation” also applies to the title of “chess grand master” due to their ever-growing ranks.
The only title in the realm of chess immune to this phenomenon, he said, is the title of “world champion”, since it is, by its very definition, exclusive to one person. Vidmar categorized chess players into four castes: Those who seize on every opportunity to take an enemy piece, he called “brutes”. One rank up are “traders” – a less greedy type of player, who doesn’t make obvious blunders. Above “traders” are “fighters” – those able to make unintuitive trades that nevertheless lead to a better position. Then, at the very top of this pyramid, is the “philosopher” chess player, who wins by way of deep analysis and a gradual building of a position.
Vidmar probably didn’t know that a computer scientist named Claude Shannon was already in the process of building the foundation for a new contender in the field – the computer. Nor did he imagine that one day a computer would be able to defeat, with ease, not only a grand master of Vidmar’s rank, but even the world champion.
It’s difficult to know for Vidmar specifically, but for most of his contemporaries, the very thought of such a thing being possible was too grotesque to even contemplate. Now, more than half a century later, we can append two new, higher castes of player to his list. One being a chess playing algorithm such as Komodo that achieves an Elo ranking of 3400. (For comparison, no human chess champion has ever even attained an Elo of more than 3000.) The other, as of yet nonexistent caste, but the idea of which is at least a century old, is the so-called “absolute” chess player – one that knows the optimal move for any given position. Interestingly, this might seem familiar, since it has always been the minimal standard for solving chess problems. When it comes to competitive chess however, even for computers, we are not there yet; nor is it certain that we ever will be, as this might be too difficult a feat even for AI – or perhaps not, we don’t yet know.
In fact, the notion of the absolute best move is already creeping into both machine and human chess. For instance, a program for analysing chess games already reports: “mate in 4 missed”. It assumes a series of four optimal moves, which inevitably lead to a checkmate, just as one might find in a chess problem. Any deviation on the part of the losing side would only mean a quicker defeat. The emergence of an absolute chess player would mean a conversion of all chess into a series of chess problems. “Black to move wins in 12 moves”, or “White draws in five moves”. An absolute chess player would come up with a similar statement for the starting chess position. Is it a victory for white, or perhaps a draw. Of course, all this under the presupposition of an absolute player on either side of the board.
Another noteworthy thing came about as a result of this chess-AI affair – centaur chess. Centaur chess is when two people play over the internet, each with an AI. Each side can also consist of many people and programs with only the hardware having to be of equal strength. The promoter of this kind of chess competition is the world champion famously defeated by an AI – Gary Kasparov. This tells us a lot about him – he really was and is in it for the chess. (As is generally known, centaurs were mythological horses with torsos of men in place of a head. Allegedly the idea of such a creature was the product of the terror of those who faced cavalry for the first time.) Thence the name – a symbiosis of the chess knight (represented by a piece shaped as a horse) symbolizing the digital algorithm and the human. The chess that takes place in this arena is played at a hitherto unimagined level. Vidmar’s four castes can but look on in silence.
The idea of concept inflation remains very relevant, especially in the field of AI. The term AI itself has become quite diluted – even more so than the title of “chess grand master” was in Vidmar’s time. Nowadays, software companies seem to stick the AI label to every product they come up with as a matter of fashion, and mostly without a convincing technical reason. That is unless we choose to class every calculator as a type of (artificial) intelligence, as they did in the pre calculator age, where the mastery of arithmetic was what, with some exaggeration, separated man from machines and animals. If we take this elevation of arithmetic seriously, then any number crunching is in fac, (artificial) intelligence. Then also is any intelligence, human or otherwise, just a front-end for the computing going on in the background – a view that some of us defend with all our heart. Nevertheless, most who flaunt their AI feathers do so without justification.
In the half a century that has passed since Vidmar’s hay-day, the field of chess has seen enormous breakthroughs, for which the better part of the credit goes not to chess players, but to AI researchers. When Deep Blue defeated Kasparov it was still in part thanks to chess experts who helped build the database of optimal chess strategies and tactics. However, in the modern, much stronger algorithms this is no longer the case – their authors are no more than mediocre chess players. In some way, for the past two decades, we have borne witness to the decline of the expert system paradigm. Not only in chess, but everywhere that the term “AI” is not used in vain.
The trend steepens even more when it comes to oriental chess – the game “Go”, which has long been a synonym for the immaturity or even the immaturitability of AI. In Go most of the advances were the work of machine learning, or, more specifically, artificial neural networks, and no longer manual, as it was in the nascency of AI chess. With Go, the better part of the code was automatically generated, albeit in the cryptic language of numeric weights and virtual synapses between virtual neurons in a virtual cortex. AlphaGo cannot yet explain to humans why it makes the moves that it does. Yet it makes them, and the moves it makes, defeat, practically without fail, any of the Go experts. From its début a year ago and until today, AlphaGo is becoming stronger every day. Well, at least as long as Google deigns to host it on its millions of computers. The improvements, as they were and as they will be, are not the work of some hypothetical grand masters and programmers, trying to encode knowledge into a computer language, as they once did – rather, AlphaGo plays against copies of itself, thereby elevating theories of playing into as of yet unimagined heights.
This shocks even the tech savvy people of the year 2017. Perhaps to the degree that it would have shaken Vidmar had he known that the intentions of Shannon (and Turing before him) would one day come to fruition – perhaps more so. Indeed, Vidmar might have been too self-critical – the grand masters of his time and those that came after, and especially many of their games, were undoubtedly worthy of going down in history. It is thus also today in the field of AI. Most of it is pretentious and unnoteworthy, yet more and more pearls of brilliance are beginning to shine through. It is these pearls that we intend to discuss and write about on this page.
The entries will be bilingual (Slovenian and English), and will chiefly pertain to concrete feats of AI.