by Angela Guess
Emma Brunskill recently wrote in The New York Times, “Artificial intelligence research is making it easier to automate decisions from data. But though AlphaGo’s success is impressive, ultimately it is winning in a game, and that can only be extrapolated so far when it comes to real world problems. AlphaGo leverages three key insights to achieve unparalleled performance: It uses deep neural networks to make decisions directly based on the board configuration; it uses extensive prior data about human-vs-human games to bootstrap its learning; and it uses reinforcement learning to improve itself based on simulated game play. These incredible abilities will undoubtedly lead to substantial benefits for other non-game applications, such as robotics.”
Brunskill goes on, “But for AlphaGo, which operates in a game, there is no real-world cost to losing as the software learns to improve. In contrast, in many applications involving people, decisions made during learning have real ramifications. A teacher-in-training or medical intern, for example, may be faced with decisions that directly affect students’ learning or patients’ quality of life. It is very hard to build good simulators of human behavior that can test new algorithms in such settings. There is a huge need for methods that can carefully explore new strategies with real people, and not just perform random exploration.”
photo credit: Flickr