Facebook continues its efforts to create artificial intelligence capable of outclassing all humans at the ancient Chinese strategy board game Go.
The social media company recently published a research paper showcasing the progress it made with the DarkForest bots, which use a synergy of methods to be the best Go players available.
Yuandong Tian and Yan Zhu, AI researchers at Facebook, explain how the computer program behaves in the abstract of the paper.
"Against human players, [darkfores2 achieves] a stable 3d level on KGS Go Server as a ranked bot," the duo points out [pdf].
This is a visible improvement over the predicted 4k-5k ranks for DCNN that Clark & Storkey (2015) reported after studying matches against other machine players.
Yuandong and Yan make it clear that the new Monte Carlo Tree Search effectively unleashed the AI's potential at playing Go. MCTS takes the tree search methods commonly deployed in computer chess programs and takes them to the next level, by "randomizing" them.
To put it in perspective, a game of chess allows for 35 legal moves per turn while a game of Go lets players make 250. The average Go game lasts for 150 turns, while a chess match stops after about 80 turns.
The point is, to have an algorithm that solves Go exhaustively would be madness, so researchers invented the MCTS. With the added tree search, darkfores2 became known as darkfmcts3. The two researchers show in the abstract that the new AI's success in the KGS Go competitions is unmatched.
"With 5,000 rollouts, it beats Pachi with 10k rollouts in all 250 games; with 75k rollouts it achieves a stable 5d level in KGS server, [...] with 110k rollouts, it won the 3rd place in January KGS Go Tournament," the abstract reads.
Facebook might be involved in the development of the Go software for months now, but the ancient game has been the subject of interest for programming experts for decades. Global competitors, such as Google and Baidu, put a lot of resources into AI research during the past few years.
Google, for example, also has an advanced Go-playing AI computer that uses machine learning, pattern recognition, planning and problem solving to defeat its human opponents. In October 2015, Google's Go playing computer, dubbed DeepMind, defeated the European Go champion, Fan Hui, in five out of five matches.
In March this year, DeepMind will meet the world champion Lee Sedol. Lee, a 9-dan professional player, is the unrivaled global master of Go for the last several decades.
Facebook seems highly motivated to catch up with the competition, as it made its AI's hardware designs public, alongside releasing the code behind DarkForest an open-source. What's more, the social media company did some heavy recruiting to strengthen its team of AI engineers.
"We think that science progresses faster when research groups exchange ideas quickly and build on top of each other's work," says AI Research head Yann LeCum.
Other notable examples of computers that beat humans at games are Watson winning on Jeopardy and chess computers downing world champions. Not sooner than 19 years ago, Deep Blue became the first computer system that defeated a reigning world champion in a chess match. It should be noted that the 1997 chess game happened under standard chess tournament time controls.
Facebook takes notes from the neural nets and reinforcement learning that were popular in the 1990s and deploys a much more sophisticated approach to them in its Go playing AI.
"We're getting close," says Facebook's Mark Zuckerberg.
Facebook will use the academic research into AI to massively upgrade its services, such as the upcoming and long expected text-based personal AI assistant, named M.
At the same time, DarkForest still has some way to go before matching the performances of DeepMind. So far, DarkForest bots were able to swipe the floor with amateur players who use the KGS Go Server. When Facebook's Go-playing AI will beat its first prominent human opponent, we will let you know in a heartbeat.
"This is an exciting time to be working on AI," LeCun concludes.