Take a brief trip down through the archives of the Nexus, back to the times when we first ventured on our mighty quest, and you will stumble across a few posts covering the idea of artificial intelligence. In particular, these were in reference to the AlphaGo program that, at the time, had become known and relevant through beating the Go world champion, Lee Sedol. Although little for popular consumption has surfaced since then, today I am happy to announce that I am setting the AI ball rolling back into action, with recent news concerning the arrival of an even more adept sibling…
At the time, the feat of AlphaGo astonished spectators from all over the world. Even with the infamous victory for the machines in 1997 with DeepBlue against Kasparov and the advanced computational powers of the modern technological age, the game of Go was believed to be impenetrable given the greater complexity in move combinations and difficulty in analysing the quality of a given board position in comparison to chess. Humans had the advantage of having studied the game meticulously over many generations, mastering it almost as an artform. However, the surprisingly innovative playstyle of AlphaGo came as a shock to professional players, who had expected a mundane, robotic approach from the hunk of metal. This rather incredible achievement was essentially brought about by feeding the machine with millions of moves made by professionals, to the point where it would be able to predict future moves with a high enough accuracy.
Going one step further, the team at Google improved the algorithm after this success so that it was based purely on self-play reinforcement learning alone, not requiring any human supervision, bar rules of the game. It started off with utterly random moves, but over several iterations was gradually able to anticipate its own moves and how they would affect a game’s outcome – in effect, it became its own teacher. During its training it even discovered a previously unknown series of moves which it found a preference for, which once again shows the creativity of the algorithm and the potential for artificial intelligence. This novel program, named AlphaGo Zero, thrashed its predecessor in a total wipeout, 100 games to 0.
Having broken free of the constraints and limitations of human knowledge, we come a step closer to a tabula rasa artificial intelligence, one able to reason and act with no built-in knowledge base of their environment, a true ‘blank slate’. This breakthrough can have a vast positive impact on society, aiding in important structural-related problems such as protein folding and the search for revolutionary new materials.
Our AI story doesn’t end here, far from it. In my next post I will show you yet further developments that have been made to this algorithm and its achievements, but bear in mind that everything is not always as it seems. I am very intrigued to see where our advancements in artificial intelligence will take us, and hope that you are equally so.