Innovation > Innovation
GOOGLE DEEPMIND, London / GOOGLE / 2016
Awards:
Overview
Credits
CampaignDescription
So far computer Go programs have only been able to play at an amateur level. During research, AlphaGo won over 99% of games against the strongest other Go programs. But the first test was to beat a professional human player, Fan Hui, the European Champion. AlphaGo won 5-0, the first time a Go program as ever beaten a professional player.
The results were published on the cover of Nature in February 2016, one of the most prestigious science journals in the world. The research community was flummoxed with words like “historic”, “breakthrough”, “milestone” the key descriptors.
The ultimate challenge was then to play Lee Sedol, the best Go player of the past decade in his hometown of Seoul, South Korea. Google DeepMind arranged a five match tournament attended by dignitaries, press and photographers and livestreamed to millions around the world. The underdog, AlphaGo, won the series 4-1, an AI milestone.
Execution
Traditional AI methods—which construct a search tree to calculate all possible positions—can’t scale to beating Go experts as there are too many possible positions.
DeepMind instead took inspiration from how the brain works and the neural networks that help humans make decisions. AlphaGo combined an advanced search tree with two neural networks. The “policy network” predicts the next move, and is used to narrow the search to consider only the moves most likely to lead to a win. The “value network”, is used to reduce the depth of the search tree and estimate the winner in each position.
AlphaGo was then trained on 30 million human expert moves. But the goal was to beat the best human players, not just mimic them. So AlphaGo learned to discover new strategies by playing thousands of games between its neural networks, and adjusting the connections using a trial-and-error process known as reinforcement learning.
Outcome
A milestone in Artificial Intelligence research has been reached.
280 million viewers tuned in on YouTube and on TV around the world; 35,000 news articles were published and the tournament generated a surge of interest in Go as people were inspired to take up the ancient game. Commentators, famous scientists and other professional players commented on the many unprecedented, creative, and even “beautiful” moves that AlphaGo made.
Unlike previous computer programs like Deep Blue that have been trained to play popular games, AlphaGo was not built specifically to play Go. So the hope is that in the future, these methods could be extended to help society with other problems, from making our phones smarter, to helping scientists tackle some of the world’s biggest challenges from climate modelling to complex disease analysis.
Relevancy
From automatic translation to fighting disease, the promise of artificial intelligence (AI) is to create smarter programs to help society solve world problems.
One outstanding grand challenge for AI researchers has been the game of Go. A game so complex that traditional AI methods are not enough.
Google DeepMind built AlphaGo, a computer Go program, which did just that by beating the best Go player of the past decade, Lee Sedol. A feat previously believed to be a decade away.
This success encourages that in future, similar methods can be used to help society tackle other real world challenges.
Synopsis
Games have long been a popular testing ground for developing computer programs that can tackle problems in similar ways to humans.The first game mastered by a computer was noughts and crosses in 1952. Then checkers. In 1997 Deep Blue famously beat Garry Kasparov at chess. The list goes on.
Then we have Go. A game that has been around for 2500 years but due to its sheer complexity, computer mastery was as yet unsuccessful. Two years ago, Google DeepMind took on the challenge.
Unlike Deep Blue and chess, the complexity of Go means it is impossible for computers to win by calculating every possible move. In Go there are more possible positions than atoms in the universe. Instead, the program would have to make judgement calls, or use what we refer to as “intuition”. An essential skill to help humans make decisions but one that is incredibly difficult for computers.
More Entries from Innovative Technology in Innovation
24 items
More Entries from GOOGLE DEEPMIND
1 items