So far, we have been trying to develop an AI that can beat human players in board games. But the race has now struck other AI players in their own games. DeepMind by Alphabet Inc. has developed an AI called AlphaZero that can learn and master games like chess, go and shogi without human intervention.
In an article published in The Journal Science The DeepMind team notes that AlphaZero is an improved version of its famous AlphaGo engine. After feeding the principles of chess, shogi and go, AlphaZero needed nine hours, 1
So it went:
- Chess: AlphaZero won 155 games, lost only six and pulled out the remainder of 1000 against Chess champion StockFish
- Shogi: DeepMind's AI surpassed the world championship software Elmo in 91.2% games.
- Go : AlphaZero beat AlphaGo in 61% of games.
AlphaZero used a search method called the Monte Carlo Search Tree (MCST) to determine the next move. This method gave her an advantage over her competitors.
In an editorial in the Science magazine, world chess champion Garry Kasparov praised the playing style of AlphaZero:
(…) I admit that I was pleased that AlphaZero had a dynamic, open style just like my own The usual wisdom was that machines with endless drying maneuvers would reach perfection, which usually led to drawn games. In my observation, AlphaZero prioritizes piece activity versus material and prefers positions that are risky and aggressive to my eye.
The secret of DeepMind's rapid learning and expertise is Google's 5000 Tensor Processing Units (TPU). To put it in context, a single TPU can process 100 million photos from Google Photos in one day, so AlphaZero benefits from a number of such.
Earlier this year, we saw that Open AI bots beat Dota 2 players. Which game will AIs master next? Any guesses?