Google’s AlphaGo AI smashing humanity’s best? Amazing stuff. Unfortunately, if you wanted to learn how the AI managed to play so well, the only option was to wade through research papers that are, frankly, not the easiest things to get through. What you need is an expert to break everything down, which is exactly what AI and autonomous robots engineer Aman Agarwal did yesterday.
In an extensive Medium post, Agarwal takes a look at Google’s rather dense AlphaGo paper and explains it in much, much simpler terms. You don’t even have to know how to play Go.
For example, here’s how Agarwal describes AlphaGo’s use of neural networks to narrow down the best moves:
So in this research, DeepMind used neural networks to do both of these tasks … They trained a “policy neural network” to decide which are the most sensible moves in a particular board position (so it’s like following an intuitive strategy to pick moves from any position). And they trained a “value neural network” to estimate how advantageous a particular board arrangement is for the player (or in other words, how likely you are to win the game from this position).
He goes on to say that by training these networks, the AI “was able to play well against state-of-the-art Go playing programs that other researchers had built before”. And by refining the techniques used by AlphaGo, it would go on to beat the game’s flesh-and-blood champions.
It’s difficult to summarise Agarwal’s work in just a few paragraphs; I’d recommend reading his full article if you have any a small interest in artificial intelligence.