AI and chess: How human brains trained machines to beat human brains

Controversies in chess can be difficult to imagine: surely nothing can go wrong in a sport involving two brilliant minds moving pieces around on a small square?

Yet, last September, Magnus Carlsen dropped out of the Sinquefield Cup after a defeat against Hans Niemann and resigned after one move in an online match against the same opposition. The chess fraternity interpreted it as Carlsen’s protest against Niemann’s cheating, something the latter subsequently admitted to.

Amidst bizarre subsequent rumours around anal beads, the scandal re-emphasised one thing: even the best players can cheat at chess.

This is not a statement about morality but the logical possibility of cheating at chess, which is not the same as feigning an injury in football to claim a foul.

Chess does not allow such options. The only way Grandmasters can cheat at chess is by taking external help from players superior to them.

But who are can play better than the best in the world? Surely not… human beings?

Can AI beat a human mind?

Machines can easily beat us at physical sports. A car can outrun Usain Bolt. Devices can be made to bowl at 200kph for hours without tiring – a barrage no human is capable of facing.

Chess is different, for the machine has to beat the human brain.

The lure of chess machines dates back centuries. The early ones were mechanical devices, which either had top chess players concealed inside or were operated by a human from.

While bona-fide machines, these innovations had little to do with artificial intelligence.

Leonardo Torres Quevedo’s 1912 creation El Ajedrecista, however, could use a king and a rook in an endgame to checkmate within a few moves. Fewer pieces, simpler situations, but undoubtedly autonomous. While most beginner-level human players can do this, it was a reasonably complex machine given the era.

The early days of AI

Alan Turing (one of the Enigma fame) and his friend, the economist and mathematician David Champernowne developed the Turochamp aka the Paper Machine. It could play basic games against beginners and solve mate-in-two problems. Turning published the program in 1951.

Programmers of the era had to encounter computers with slow processing speed. They were forced to use ‘selective, knowledge-based algorithms’. Which is to say, the early chess programs ‘thought’ like humans. They chose from what human wisdom perceived as ‘good moves’ largely because the archaic computers could do little more.

Some programmers tried to work around these handicaps. Los Alamos chess (a variant played on a 6×6 board without complications like castling, en passant, etc) was written in 1956.

But chess programs largely continued to imitate their creators, and as that happened, the big question loomed. If and when computers worked faster than human brains, would they also work better than them at chess?

A quicker computer could, after all, apply the ‘brute force’ method – of enumerating every possible option (about a billion every move) to identify the best move.

Of bets and battles

Kotok-McCarthy, the first computer program to play ‘convincing’ chess, came in 1962.

As integrated circuits arrived, computers became quicker and, as a logical outcome, better at chess.

The Association for Computing Machinery hosted the North American Computer Chess Championship from 1970. The World Computer Chess Championship began in 1974.

The software program Chess, developed by the Northwestern University, was considered one of the best of its tine. So much that Chess 4.6 was crowned the world champion in 1977.

In 1978, English International Master David Levy beat Chess 4.7 by a 4.5-1.5 margin to claim a wager from 1968: he had bet that no computer program would be able to beat him in 10 years.

Despite being vindicated, Levy admitted that his opponent much stronger than he had thought possible when he started the bet.

A decade later, in 1989, Deep Thought, developed at the Carnegie Mellon University and later at IBM, would trounce Levy 4-0. Just the previous year, Deep Thought had defeated Bent Larsen to become the first known program to trump a grandmaster.

Before that, in 1983, Belle had become the first program to attain Master level. Kasparov realised that Grandmaster, the obvious logical next step, was only a matter of time: “The writing was on the wall,” he famously said, “Even if many of us would stay in denial for another decade.”

The best, and beyond

In 1995, Fritz – a German chess program developed by Frans Morsch – defeated Deep Blue (the erstwhile Deep Thought) to win the World Computer Chess Championship – the first triumph of a consumer-level microcomputer over a mainframe.

In 1996, Deep Blue took on – and lost to – Kasparov by a 2-4 margin. But it was the 1997 contest that got even those outside the chess fraternity interested.

It came down to the decider after the score was tied 2.5-2.5. Subsequent experts spotted three errors by Kasparov in Deep Blue’s 19-move devastation, the shortest of the twelve matches.

The duel was not without controversy: Kasparov alleged IBM of cheating, of taking a human grandmaster’s help (a quarter of a century later, Niemann would be accused of exactly the opposite!). But everything got buried under the enormity of the first instance of a program defeating a world champion.

In the early 21st century, Deep Fritz drew with Vladimir Kramnik and X3D Fritz with Kasparov. In 2009, Pocket Fritz, running on a mobile phone, won the Copa Mercosur in Buenos Aires.

The obvious question that arose: If a mobile phone could do that, how good were the best computers or laptops?

And, pertinently, what was left to achieve?

The solution was inside

The answer came in 2017, from outside chess. AlphaGo, designed by DeepMind to play go, a game more complicated than chess and notoriously difficult, beat itself by “eschewing embedded human knowledge and teaching itself to play better”.

Clearly there were methods superior to ‘brute force’ in go.

Taking a cue, DeepMind launched AlphaZero to ace chess, shogi, and go. The program does not need to analyse existing matches played by humans or computers. Instead, it teaches itself based on its own past mistakes by playing “more games against itself in a few hours than have ever been recorded in human history”.

Contrast this with Maia Chess, the antitheses of an AlphaZero. Maia plays the most human – not necessarily the best – response to every situation by “matching moves played by human players in online games over 50 per cent of the time.”

While AlphaZero strives to increase the already vast chasm between AI and the human brain, Maia wants to bridge the gap, to become the most ‘human’ chess program.

The future

One can understand how programs like AlphaZero can induce paranoia in us. Machines improving by teaching themselves ad infinitum is not a threat to us in chess, but what about life? Should we be wary of the takeover by machines that science fiction has warned us for so long?

After centuries of imagining machines as tools intellectually inferior to us, it is time we accepted that computers are not only more skilled than us, but they will keep getting even better.

The path to success will undoubtedly involve humans treating them as ‘consultants’ or ‘experts’ and making the most of their unimaginably superior skills.