Winning by Insights... and Other Means

Here’s how Deep Blue, the computer that famously defeated (then) world champion Garry Kasparov in the pre-Big Data age “learnt” to play, writes Steven Strogatz:

“Principles, which have been refined over decades of human grandmaster experience, are programmed into the engines as complex evaluation functions that indicate what to seek in a position and what to avoid”

Sure, it was insanely fast, but was “utterly lacking insight”.

Today’s chess engines “learn” by playing and looking at games, “discovering” for themselves patterns. It’s how all of machine learning works: the more and better the data you throw at it in the learning phase, the better it gets. With machine learning techniques, computers often appear to identify insights never known to humans! Take this description of AlphaZero, Google’s machine learning algorithm world champion at chess, shogi (Japanese chess) and Go. It crushed its “programmed” learning counterpart, Stockfish by:

“thinking smarter, not faster; it examined only 60 thousand positions a second, compared to 60 million for Stockfish. It was wiser, knowing what to think about and what to ignore.”

Kasparov himself hailed AlphaZero saying it “reflects the truth” rather than the “priorities and prejudices of programmers”.

Machine learning is increasingly proving better and faster in other areas like medical diagnosis as well. But, and this is becoming a problem now:

“The algorithms can’t articulate what they’re thinking. We don’t know why they work, so we don’t know if they can be trusted.”

But why bother about how it works as long as it works? Aha. Take this AI that applied machine learning on an old arcade (video) game called Q’bert. Soon it was getting “impossibly high scores”, writes Charlie Osborne. Since this was a video game, it was easy for researchers to see how it was doing this. The surprising answer? The AI had found a bug in the game, something it then exploited to rake up a million points!

Does it even make sense to say that the AI cheated? Now we’re entering philosophy, so I’ll take a diversion into something my 7 yo daughter did recently. I was trying to introduce her to computer programming via this made-for-kids language called Scratch. I talked her through the program code for a sample game, and then let her play the game. I soon found her getting “impossibly high scores”. How? She’d gone and changed the program code to increase the score by 10 (instead of the original 1) every time she did what was expected. I was stuck in two minds. Should I be happy that she was programming? Or unhappy with her cheating?

I still haven’t made up my mind.