江苏龙网

  1. 首页 > 生活 > >

AlphaGo Zero 能够取得突破的关键瓶颈在哪里( 二 )


I guess it\u0026#39;s a question of people and resources and priorities! If we\u0026#39;d run for 3 months, I guess you might still be wondering what would happen after, say, 6 months :)
We have stopped active research into making AlphaGo stronger. But it\u0026#39;s still there as a research test-bed for DeepMinders to experiment with new ideas and algorithms.
(1) I believe these "bad" moves of AlphaGo are only bad from a perspective of maximising score, as a human would play. But if the lower scoring move leads to a sure win - is it really bad?
(2) AlphaGo has learned plenty of human joseki and also its own joseki, indeed human pro players now sometimes play AlphaGo joseki :)

AlphaGo is retired! That means the people and hardware resources have moved onto other projects on the long, winding road to AI :)

The creators of the documentary are planning a digital release in the next few months on platforms where you can buy and rent movies, such as Google Play Store, iTunes, YouTube Movies. They’re also currently exploring a release on a streaming service too.

Life at DeepMind is great :) Not a recruitment plug - but I feel actually quite lucky and privileged to be here doing what I love every day. Lots of (sometimes too many! :)) cool projects to get involved in.
We\u0026#39;ve been lucky enough to have many great people work on AlphaGo - you can get an idea of the contributors by looking at the respective author lists - also there is a very brief outline of contributions in the respective Nature papers.
Facebook focused more on supervised learning, producing one of the strongest programs at that time. We chose to focus more on reinforcement learning, as we believed it would ultimately take us beyond human knowledge. Our recent results actually show that a supervised-only approach can achieve a surprisingly high performance - but that reinforcement learning was absolutely key to progressing far beyond human levels.
Actually, the representation would probably work well with other choices than 8 planes! But we use a stacked history of observations for three reasons: 1. it is consistent with common input representations in other domains (e.g. Atari), 2. we need some history to represent ko, 3. it is useful to have some history to have an idea of where the opponent played recently - these can act as a kind of attention mechanism (i.e. focus on where my opponent thinks is important). The 17th plane is necessary to know which colour we are playing - important because of the komi rule.
We just asked Fan Hui about this position. He says AlphaGo would solve the problem, but the more interesting question would be if AlphaGo found the book answer, or another solution that no one has ever imagined. That\u0026#39;s the kind of thing which we have seen with so many moves in AlphaGo’s play!
AlphaGo Zero uses a quite different approach to deep RL than typical (model-free) algorithms such as policy gradient or Q-learning. By using AlphaGo search we massively improve the policy and self-play outcomes - and then we apply simple, gradient based updates to train the next policy + value network. This appears to be much more stable than incremental, gradient-based policy improvements that can potentially forget previous improvements.


推荐阅读


上一篇：微信里面的360度全景是咋做到的

下一篇：鲍鱼|天冷了，这汤多煲给家人喝，鲜香甘甜，常喝增强免疫力，对身体好