Blog: The Age of Alpha Go “Zero”
Not long ago, witnessed by tens of millions of people, the South Korean professional Go player of 9 dan rank Lee Sedol lost 4–1 to Aplha Go, the A.I. Go playing program created by Deep Mind.
Go has always been considered the most complicated chess game, the amount of all possible moves in a single game are more than the stars in the universe. Unlike the previous A.I. chess player Deep Blue, AlphaGo can’t win this challenge by simple consider all the moves, instead, it must learn how the human “thinks”.
Created by Dr. Huang, Alpha Go is trained by human professional Go player. By showing it thousands of games played by professional players, Alpha Go is able to study the pattern of how the players moves. From these moves Alpha Go could learn how the professional players “think” under certain situations and even follows “intuition” when facing unseen moves to make the correct decision.
Well if you think Alpha Go is the best Go player we could get, wait until you met the new A.I. that beat Alpha Go like nothing. Shortly after Alpha Go defeated Lee Sedol, DeepMind, the parent company of Alpha Go, published their new A.I. algorithm — Alpha Go Zero, and claimed that it has defeated Alpha Go more than a thousand times. The main difference between Go and Zero is that Alpha Go Zero does not refer to any human knowledge, it learn how to win by playing the game against itself million and million times.
What Alpha Go Zero has achieved and the meaning behind are in fact really shocking. The wisdom humans have accumulated in Go for thousands of years now seems to be like an “obstacle” between A.I. and perfection. The best solutions we got are all bounded in what human brains can handle. Without constraining A.I. with these boundaries, A.I. is able to found a more better solution that human brains could never imagine of.
Well, what could this lead us to? In the field of medication, Penicillin has always been one of the most effective antibiotics. Considering the application on Go, what if we define the game as who could kill the most bacteria, provide all possible ingredients and chemicals that may possibly be used, and let the A.I. do its work. Instead of playing the game our self like what humans did in the past centuries, we only define the goal of the game. This not only save human efforts on try and errors, but could also achieved a better solution that human brains could never achieved.
Aside from medication, stock trading is also an example of finding solutions by A.I. Many bank has already been recruiting A.I. experts and data scientist to analyze the market for them. These experts don’t really have solid economic background like those in business schools, but they are capable of creating A.I. that studies the techniques in trading and make decisions for them. Maybe after years, we won’t have legendary like Warren Buffett in stock market. Instead, the market will be full of AI traders competing one another.
Aside from those examples, there actually many applications that shows the capability of A.I. breaking limitation in human knowledge. The picture above is the optimized car model created by A.I. What human did only in this design is defining the way to “win the game”, which is remain the safety of the driver during crash and having the lowest production cost. It is almost impossible for human designer to create structures like this. This is only just one of the examples what A.I. could do. Imagine other application domain like architecture, agriculture, auto-pilot, aerospace ……
Alpha Go shows that A.I. could learn from human, while Alpha Go Zero shows that A.I. could learn better without human. In the near future, human will be solving problems by telling A.I. what is the goal and making it find the best solutions. Imagine a world that we don’t need experiment thousand and thousand of times just to find a single solution. This will not only bring us breakthroughs in many domains, but also free us from the repeating experimenting labors and allow us to do more creating things. Imagine a world that human ask questions and A.I. give answers. That’s the future that we are heading.