Rise of AI in strategy: AlphaGo

Chess was an easy cake for AI. How about Go?

AI has been a topic of discussion and curiosity for a long time, but the debates to got more and more intense especially by the end of the twentieth century. The research and discussion about it started in 1950s and it was (dis)popularized by movies like The Terminator (although it described as nothing more than an invisible evil). Most times it was viewed as an entity that will transform everything in our life (both in positive and negative ways), especially in business. It is of debate how much this had happened, but it made a big headline in 1997 when a computer program named Micosoft DeepBlue defeated the world champion Garry Kasparov in chess game.

Although it was heavily researched and used in industry, AI has not been in too much debated in media since the chess defeat, until AlphaGo (owned by Google DeepMind) defeated Lee Sedol, one of the best players at Go game in the world 4-to-1 in 2016. This event was quite different from Kasparov's defeat two decades ago. As an expert system, DeepBlue was programmed by embedding the wisdom of human chess experience into a computer program by humans and it calculated potential gain and loss of each move by considering the consecutive actions by the opponent.

Compared to chess, the rules of go game is extremely simple (all stones are same and the goal is occupy more space in the board), yet mastering it (for a computer program) is extremely difficult. For a computer to calculate next 8 moves, a computer algorithm has to calculate 512 quintillion (5.12x10^20) combinations, which is essentially impractical even for the strongest combinations. An algorithm has to develop intuition of strategy to defeat a master of this game. And that's what basically what AlphaGo algorithm did. A very limited amount of simple techniques were manually embedded into the program. Then, a neural network (a.k.a deep learning ) was trained with the database of games played by human masters, which contained millions of moves. Finally, AphaGo was further trained by playing against itself using a reinforcement learning technique and reached to a level to defeat the best go players in the world.

Later, DeepMind developed AlphaGo Zero, which is an advanced go player. Unlike AlphaGo, in which certain game techniques are programmed manually, AlphaGoZero uses no input except the rules of the game. It was trained by playing against itself many times. AlphaGo Zero was made to play Against AlphaGo Lee (the version of AlphaGo that defetated Lee Sedol) and won 100-to-0 against it, proving its invincibility against any human player. Finally, DeepMind team developed AlphaZero, which mastered the games chess, go and shogi, sealing the discussion about power of AI in strategy games.

The success of AI in AlphaGo sparked different effects in communities. In the AI community, it was excitement and joy. It the go community, it was a big surprise and shock (and a little disappointment at least at first) to see their unconquerable game to be mastered by a computer and that any human player in the world can easily be defeated by a computer program.

But what the implications would be for the overall industry and public? The success of AlphaGo sparkled new discussions in (social) media, about how powerful the AI can be and which year each proficiency will be replaced by AI. Was this all true? Could we rely on AI do all the jobs (driving, teaching, law, medicine, even programming) and humans can just enjoy and entertain themselves? What would the implications be? Who would benefit and gain wealth and who would loose? What should the society and governments embrace and manage the implications? These are some of the questions of many, which are related to AI and and requires answers from many people. In my future articles, I will try to focus on its implications in science, research and (bio)medical industry.