Google's AI Beats Human Professional Player At Go
Written by Mike James   
Wednesday, 27 January 2016

Go has long been a troublesome game for AI. It was the game in which humans clung onto supremacy - until now. A neural network based system has finally beaten a really good human player.

googleresearch

 

The new system achieved a 99.8% winning rate against other Go programs and defeated the (human) European Go champion by five games to zero. Of course, this doesn't make it provably the best Go player in the world, but you can tell that something big has just happened. Computers can now beat us at Go!

The architecture of the network, called AlphaGo and designed by Google's Deep Mind team, is novel in that it uses two distinct neural networks in a traditional reinforcement learning "actor-critic" arrangement. The first network picks possible moves and the second evaluates them in terms of how much advantage the move gives in the future. This approach allowed AlphaGo to learn from 30 million human games of Go and then go on to refine its performance by playing millions of games against itself. 

deepmind

You can see how the game and AlphaGo works in the following video:

 

 

This is a more "human" approach to playing games. The standard AI approach to playing games is based on deep search to evaluate positions. Essentially chess playing programs win because they have the computational power to play the game into the future and evaluate the next move in terms of what will subsequently happen. AlphaGo makes more use of the intelligent selection of moves, which it then evaluates. This seems to be what is needed to enable a machine to win at Go.

One thing that is certain is that the well known and previously "accurate" xkcd cartoon will need to be updated:

Game AIs

 More cartoon fun at xkcd a webcomic of romance,sarcasm, math, and language

 

There are plans to evaluate AlphaGo by playing against the world's best players so watch for more news soon. 

What is interesting is that Google's Deep Mind seems to have found a way of combining the ideas from reinforcement learning and neural networks. 

 

Goplayicon

 Picture Credit: Goban1

 

More Information

Mastering the game of Go with deep neural networks and tree search

Related Articles

Google's DeepMind Learns To Play Arcade Games

Neural Networks Beat Humans 

Google DeepMind Expands 

Google Buys Unproven AI Company

Deep Learning Researchers To Work For Google

Google's Neural Networks See Even Better

 

To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on, Twitter, FacebookGoogle+ or Linkedin

 

Banner


C23 ISO Standard Is Here But You Probably Won't Read It
06/11/2024

At last ISO C23 has been published, but at $250 you probably aren't going to read it. Can we really tolerate this sort of profiteering on the work of others? This is worse than academic publishing!



OpenAI Releases Swarm
25/10/2024

OpenAI has released an experimental educational framework for exploring ergonomic, lightweight multi-agent orchestration. Swarm is managed by the OpenAI Solution team, but is not intended to be used i [ ... ]


More News

 

espbook

 

Comments




or email your comment to: comments@i-programmer.info

 

Last Updated ( Wednesday, 27 January 2016 )