Glossary
What is AlphaGo?
In short
AlphaGo is a program from the DeepMind lab that plays Go; in March 2016 it beat Lee Sedol 4:1.
History of AI · AlphaGo
MMOOZZGG.com editorsChecked: 1 min read
For a child
Go is an old board game with black and white stones. It has so many possible moves that no machine can count them all. Scientists made a program that first watched how masters play and then played a huge number of games against itself. This way it learned to feel a good move, and it beat a famous master while the whole world watched.
In plain words
Go has far more positions than chess, so you cannot check them all the way chess computers did. AlphaGo solved the task in another way. One neural network suggested which moves to consider, another judged who was winning in a position, and a search checked several continuations. First the program learned from the games of masters, then it improved by playing against itself. In the second game of the match with Lee Sedol it played move 37, which, according to DeepMind, had a 1 in 10,000 chance of appearing in a game.
For business
AlphaGo showed business two things. First, a system can find a solution that is not in human experience if it can test options by itself and get a score for the result. A similar principle is being tried in production planning, logistics and inventory management. Second, such projects are expensive. You need a simulator where the system can safely fail many times, and a metric it cannot cheat. Without a digital copy of the process, the program would learn on live customers and money, and that is not acceptable. So companies more often take ready forecasting and optimization models and keep reinforcement learning for tasks where the environment can be modeled exactly. Where this approach will pay back is assessed before development starts.
For a specialist
Silver and co-authors described the program in the paper “Mastering the game of Go with deep neural networks and tree search” (Nature, January 2016). A convolutional policy network chose candidate moves, and a value network judged the position; the policy network first learned with supervised learning on expert games, then with reinforcement learning in games against itself, the value network was trained on positions from those games, and in play Monte Carlo tree search (MCTS) combined them. In October 2015 AlphaGo beat three-time European champion Fan Hui 5:0, the first AI win over a professional at Go. In March 2016 in Seoul it beat Lee Sedol 4:1, more than 200 million people watched the match, and the program received the highest professional rank of 9 dan. AlphaGo Zero (October 2017) did without human games: it started from random moves and after three days of self-play beat the earlier version 100:0. DeepMind estimates the game space at 10 to the power of 170 board configurations.
Example
A logistics company wants to teach a system to assign trucks to orders. Before they connect it to real trips, engineers build a simulator from the order history, and the system trains there, as AlphaGo trained in games against itself.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map