GOOD NEWS CHAMPION games players: DeepMind has got bored of embarrassing you and has turned its attentions to besting other artificial intelligences instead.
While AlphaGo was named due to its affinity for ‘Go’ (kind of like christening athletes “Jennifer Javelin” or “Sean Shot Put”), DeepMind’s latest AI – AlphaZero – is a jack of at least three trades, and was able to learn and master chess, Go and shogi without any human intervention, aside from a basic summary of the rules.
If you’re wondering which game is hardest, it looks like it’s Go, by the way. While it took AlphaZero 13 days to get on top of Go, it had mastered chess and shogi in just nine and 12 hours respectively. After this training programme, AlphaZero put its lessons into practice, taking on the world’s best AIs in each game.
First up, AlphaZero took on Stockfish at chess. Yes, with those names you’d think they’d be battling it out on Xbox Live, but this was both more serious and a lot more drawn out that your average Halo match. In total, the two AIs played 1,000 games of chess, and the final score was AlphaZero 155, Stockfish 6. That’s right: the remaining 839 matches were draws, so be glad you didn’t get this series on pay per view.
In shogi, things were a lot more clearcut. AlphaZero beat Elmo – presumably not the muppet with a penchant for tickling – in 91.2 per cent of the matches.
Go was relatively tight, which isn’t hugely surprising as AlphaZero was taking on its predecessor, AlphaGo. Despite this, AlphaZero’s new learning methods paid off, and it ended up beating AlphrGo 61 per cent of the time.
One person who did dip into at least some of the 1,000 chess matches was legendary human chess champion Garry Kasparov who famously lost to Deep Blue 21 years ago, before AlphaZero was a glint in Google’s eyes. Writing an editorial in Science Magazine, Kasparov explained how impressed he was at the free-flowing chess played by the AI, despite the 839 draws.
“I admit that I was pleased to see that AlphaZero had a dynamic, open style like my own,” he wrote. “In my observation, AlphaZero prioritises piece activity over the material, preferring positions that to my eye looked risky and aggressive.”
It may look effortless, but the 13-day training programme was only made possible by some very powerful hardware. The training process used 5,000 tensor processing units. Just one of those bad boys chews through over 100 million pictures in Google Photos every day, which means that if you’re wanting to train your own AI to beat AlphaZero, you’d best start saving now. µ
Source : Inquirer