AI researchers thought that it would take decades before an Artificial Intelligence algorithm would beat a human grandmaster in the complex game of Go. But Google’s AI, AlphaGo, had just surpassed their expectations last week by beating a grandmaster in a five-match game of a Man-vs-Machine face off and winning the $1 million prize money.
AlphaGo emerged as victor against the 33-year old South Korean Go grandmaster Lee Se-Dol in a week-long battle from March 9 to March 15 in Seoul, Korea. The AI took the first three games in a decisive and convincing sweep that shocked a confident Lee, with analysts saying AlphaGo was so formidable as a Go player and that it played so well it was scary.
Lee brought out his best plays on the fourth game where he finally took home a sweet and hard-earned victory. On his winning game, Lee changed tactics and played a strong, quick and aggressive strategy early on after realizing that the longer the game pans out, the better the AI can analyze possibilities and choose more effective moves. After AlphaGo committed a mistake on move 78, it made a series of bad moves and when it computed a probability of less than a 20% win, it resigned.
DeepMind, a subsidiary of Google, designed AlphaGo’s system based on deep learning and reinforcement learning, enabling AlphaGo to do simulated games against itself with access to about 30 million moves and patterns. Lee utilized the same strategy in their fifth and final game, but AlphaGo played an almost impeccable game following its loss.
At a post-match press interview, Lee described his single win as one “so valuable and I will not trade this for anything in the world.” DeepMind’s founder Demi Hassabis praised the grandmaster for his for brilliance and fighting spirit, even after his consecutive defeats against AlphaGo. The matches would also provide a good research material to look into the AI’s weakness.
The Man-vs-Machine matches started December last year, where AlphaGo won 5-0 against European Go champion Fan Hui in Britain. AlphaGo, still only a two-year-old program, was compared to being a child prodigy that lacks experience. Experts had predicted then that AlphaGo and the 18-time world champion Lee would have equal chances of winning, hence their surprise at an almost sweep of the AI in this series.
The game of Go is more than 2,000 years old and is considered more complex than any other board games that measure human intellect, with Hassabis describing it as the Holy Grail of AI programming. For AI experts, the defeat of a human of Lee’s caliber is a breakthrough as it signals the fast and impressive progress of smart computers, far exceeding human predictions and expectations. While some specialists celebrate AlphaGo’s win as a scientific advancement, other AI experts expressed concerns that the progress might lead to developments that could go far beyond human control.
Note: The prize money was donated to UNICEF, STEM, and Go charities.