AlphaGo's Historic Win Predicted by Liulishuo Scientists

Some experts predicted Ke Jie had a better chance to win than Lee Sedol, the last human Go player to officially test the AI as he came better prepared for Google's DeepMind Go-playing machine AlphaGo. Dr. Hui Lin had a different outlook, however. "The machine has a huge advantage in calculating and processing sophisticated situations. Moreover, when heading to the later stages of the game, a human can never control their emotion like a machine does."

"Looking back on AlphaGo1.0, AlphaGo2.0 offers a more significant mastering of artificial intelligence over board games, with its unattainable amount of training," Dr. Hui Lin said. "At this point, AI's intelligence basically relies on its gigantic database. AlphaGo's input includes every manual and game record anyone could think of, so it's safe to say that it knows its opponents well enough. AI has many widely accepted qualities like ultra-speed calculation, but its advantage over conventional training can only be revealed when it is upgraded."

Dr. Lin also explained from a technical perspective why Ke Jie described AlphaGo's moves as "mysterious" during warm-up stages, "AlphaGo won because it was unpredictable. Technically, we can't say that AlphaGo is the best Go student ever 'taught' by human beings, because it could never be restricted by any human rules. Incorporating lessons from AlphaGo1.0, AlphaGo2.0 was able to learn from all manuals and records in existence. AI never stops playing by itself, as it never eats or sleeps, and eventually it develops its own strategy, regardless of human restrictions."

A graduate from the world's leading Swiss AI Lab IDSIA and instructed by the "father of artificial intelligence" Jurgen Schmidhuber, Dr. Yi Sun has always paid close attention to DeepMind co-founder Shane Legg's AI projects, who also happens to be an IDSIA alumnus. After today's game, he praised the technical improvements of this AlphaGo version in Wuzhen, compared to how the project performed back when it was still in development under the pseudonym "Master".

Dr. Yi Sun considered reinforcement learning technology as AlphaGo2.0's main highlight. As this technology was one of his research areas at PhD level, Dr. Sun explained the purpose of this concept from the inside out, "To accomplish a sophisticated goal through the interaction of an AI and unknown environments, which in this case means to make judgments on a given Go game situation. Reinforcement learning allows the AI to study every middle-game as a "value function" and every move as a "q function" from an ocean of game records. AlphaGo built a new model in this environment, based on AlphaGo1.0's interaction history. This model would go on to learn the best outcomes of each situation through numerous mistakes, and learn to make plans in the face of uncertainty. As a result, its anticipation and positioning skills also improved to the next level."

Reinforcement learning now shows strong potential in practical applications also. Liulishuo's AI English Teacher was self-developed based on reinforcement learning algorithms similar to those described by Dr. Yi Sun. Along with the supervision of its two specialists, Liulishuo's AI builds an adaptive learning system through continuous reinforcement learning and self-evolution. The result is a truly personalized and adaptive learning path for every user, providing better learning efficiency.

On May 24th both AlphaGo and Ke Jie will take one day off, and AI specialists from DeepMind and Google will discuss the future of AI as part of the Future of Go Summit being held in Wuzhen, China. Yi Wang, the CEO and co-founder of Liulishuo will also share his ideas of applying AI to education during the summit. As for Ke Jie, he still has the chance to defeat the upgraded AlphaGo during the two remaining games on May 25th and 27th.

SOURCE Liulishuo