The year 2016 was memorable in the history of machine intelligence, marking the end of one era and the start of another. Sixty years previously, in 1956, ten computer scientists, including John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, coined the term ‘Artificial Intelligence’ (AI) for a conference held at Dartmouth College. In early 2016, the last surviving founding father of AI, Minsky, passed away — perhaps an indication that the first phase of humans pursuing machine intelligence had come to an end.
Just two months after Marvin Minsky died, Google’s AI computer program AlphaGo defeated Korea’s Lee Se-dol — one of the world’s best Go players — in a landslide 4-1 victory. AlphaGo became the first robot to take down a world-class Go champion. Because Go has six to nine orders of magnitude more complexity than chess, the event set a new precedent with implications more far-reaching than the 1997 match when IBM’s Deep Blue supercomputer beat chess grandmaster Garry Kasparov.
This AlphaGo event marked the advent of a new era in machine intelligence.
The time has come when intelligent machines are dominating humans across all types of strategy-based games. The reason is that computational capacity continues to grow exponentially while that of humans only increases linearly, at best. Although a majority of people believed that the day would come sooner or later, most felt that it wouldn’t happen for many years to come.
In fact, before the game between AlphaGo and Lee, most people, including Lee himself, Nie Weiping (a Chinese Go grandmaster), and even Dr. Kai-Fu Lee (a former Google executive), did not believe that AlphaGo could win. They reached such a conclusion because they knew the daunting difficulties in playing the Go game, not because they did not understand the latest developments in today’s machine intelligence.
At the end of 2015, AlphaGo beat only the 2-dan professional Go player, Fan Hui, who was far below the highest 9-dan professional ranking. However, many expert observers did not notice that AlphaGo was improving so quickly as was proven later. In fact, prior to the match, insiders at Google already knew that AlphaGo’s capabilities had already reached the 9-dan professional level.
The reason computers can defeat humans is that machines become intelligent in a different way than humans do. Machines do not rely on logical reasoning; instead, they heavily depend on Big Data and intelligent search and pattern-matching algorithms.
AlphaGo has been trained with data from hundreds of thousands of previous matches between the world’s top Go players. This is a key reason why AlphaGo is so intelligent. Additionally, tens of thousands of servers have been used for AlphaGo to compute and train its own Go-playing model, and different versions of AlphaGo have played with each other in tens of millions of games. Together these measures ensured that AlphaGo would possess an unmatched computing capability.
In terms of computing a game strategy, AlphaGo employs two key technologies:
Turn the current state of the Go board into a winning-probability mathematical model. This model does not have any artificial rules but is entirely obtained from previous data training.
The Monte Carlo Tree Search (MCTS) is a heuristic algorithm that precisely limits searches to a minimal range so that the computer can quickly find the next best move.
The training model and algorithms used by AlphaGo are not new at all. They are actually machine learning and game tree search algorithms that have been known for decades. What Google has accomplished is having these algorithms run simultaneously over tens of thousands, or even millions of servers. The result is a dramatic improvement in the capabilities of computers to solve difficult problems. These algorithms were not designed specifically for playing Go, and many of them have proven successful in intelligent applications such as speech recognition, machine translation, image recognition, and Big Data-driven healthcare.
Although AlphaGo’s training has used tens of thousands of servers, only dozens of servers (with more than 1,000 cores and over 100 GPUs) were employed for the games with Lee. Compared to chess, Go requires a much larger search space, and even though the computing capabilities of AlphaGo are not much greater than those of IBM’s Deep Blue, the AlphaGo search algorithms are substantially improved. With such algorithms, AlphaGo can accurately focus within targeted search spaces to calculate the next best move in a very short time.
The ultimate goal of Google in developing AlphaGo is not to prove that computers can play Go better than humans but, instead, it is to develop a machine-learning tool that allows computers to solve intelligence problems. The match between AlphaGo and Lee is only a test of how far machine intelligence has come, and there is no question that Google’s success would be impossible without the contributions from players such as Lee. AlphaGo’s victory marks a new level of machine intelligence and a victory for mankind.
AlphaGo’s success shows that machine intelligence has reached a new level but, most important, it signifies that computers are solving a wider range of intelligence problems. Today, computers have begun to accomplish many tasks that used to be done by humans, such as conducting medical diagnoses, reading and processing documents, answering questions automatically, writing press releases, and driving cars.
The AlphaGo victory is a shock to people who were not familiar with machine intelligence. The worry is that machines will one day control humans. This fear is actually unfounded because the essence of AlphaGo is a software program developed by computer scientists. Machines do not control humans, but rather the people who make smart machines can and do control the machines.
Science and technology are always playing an active and revolutionary role in human progress, and their development cannot be stopped. What we can do is face reality and seize opportunities arising from the intelligence revolution, rather than try to evade, deny, or stop it.
The future of society belongs to creative individuals, including computer scientists, rather than those who master single, specific skills that involve repetitive work.