World Go champion Lee Se-dol has beaten AlphaGo, an AI program developed by Google’s DeepMind unit this weekend, though he still trails the program 3-1 in the series.
Google’s publicity stunt highlights the progress which has been made in the world of artificial intelligence and machine learning, as commentators predicted a run-away victory for Se-dol.
DeepMind founder Demis Hassabis commented on Twitter “Lee Sedol is playing brilliantly! #AlphaGo thought it was doing well, but got confused on move 87. We are in trouble now…” allowing Se-dol to win the fourth game in the five game series. While the stunt demonstrates the potential of machine learning, Se-dol’s consolation victory proves that the technology is still capable of making mistakes.
The complexity of the game presented a number of problems for the DeepMind team, as traditional machine learning techniques would not enable the program to be successful. Traditional AI methods, which construct a search tree over all possible positions, would have required too much compute power due to the vast number of permutations within the game. The game is played primarily through intuition and feel, presenting a complex challenge for AI researchers.
The DeepMind team created a program that combined an advanced tree search with deep neural network, which enabled the program to play thousands of games with itself. The games allowed the machine to readjust its behaviour, a technique called reinforcement learning, to improve its performance day by day. This technique allows the machine to play human opponents in its own right, as opposed to mimic other players which it has studied. Commentators who has watched all four games have repeatedly questioned whether some of the moves put forward by AlphaGo were mistakes or simply unconventional strategies devised by the reinforcement learning technique.
Although the AlphaGo program demonstrates progress as well as an alternative means to build machine learning techniques, the defeat highlights that AI is still fallible; there is still some way to go before AI will become the norm in the business world.
In other AI news Microsoft has also launched its own publicity stunt, though Minecraft. The AIX platform allows computer scientists to use the world of Minecraft as a test bed to improve their own artificial intelligence projects. The platform is currently available to a small number of academic researchers, though it will be available via an open-source licence during 2016.
Minecraft appeals to the mass market due to the endless possibilities offered to the users, however the open-ended nature of the game also lends itself to artificial intelligence researchers. From searching an unknown environment, to building structures, the platform offers researchers an open playing field to build custom scenarios and challenges for an acritical intelligence offering.
Aside from the limitless environment, Minecraft also offers a cheaper alternative for researchers. In a real world environment, researcher may deploy a robot in the field though any challenges may cause damage to the robot itself. For example, should the robot not be able to navigate around a ditch, this could result in costly repairs or even replacing the robot entirely. Falling into a ditch in Minecraft simply results in restarting the game and the experiment.
“Minecraft is the perfect platform for this kind of research because it’s this very open world,” said Katja Hofmann, lead researcher at the Machine Learning and Perception group at Microsoft Research Cambridge. “You can do survival mode, you can do ‘build battles’ with your friends, you can do courses, you can implement our own games. This is really exciting for artificial intelligence because it allows us to create games that stretch beyond current abilities.”
One of the main challenges the Microsoft team are aiming to address is the process of learning and addressing problems. Scientists have become very efficient at teaching machines to do specific tasks, though decision making in new situations is the next step in the journey. This “General Intelligence” is more similar to the complex manner in which humans learn and make decisions every day. “A computer algorithm may be able to take one task and do it as well or even better than an average adult, but it can’t compete with how an infant is taking in all sorts of inputs – light, smell, touch, sound, discomfort – and learning that if you cry chances are good that Mom will feed you,” Microsoft highlighted in its blog.