TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

By sparring with AlphaGo, researchers are learning how an algorithm thinks

19 点作者 chwolfe超过 8 年前

3 条评论

alanfalcon超过 8 年前
So I&#x27;ve been following the english Go community as an outsider ever since the first AlphaGo matches, and it&#x27;s been great and fascinating to see what an impact AlphaGo has had across all levels of play.<p>Fan Hui has the advantage of being able to peek behind the curtain, but other pro players have still been studying every public AlphaGo match to try to find what sequences and types of moves it prefers, and which well-known and commonly accepted sequences it completely avoids. The problem, of course, is that if you simply imitate without understanding then you may end up doing stupid things. In fact, even AlphaGo did this against Lee Sedol whenever it played out forcing moves which only served to limit its options and point potential—in some cases strictly losing itself points—all because nobody teaches AlphaGo that when humans play these moves it&#x27;s often because they&#x27;re into overtime and simply have to play a move before time runs out to avoid losing on time, and it gives the human more time to read out a complicated sequence elsewhere on the board.<p>So while it&#x27;s interesting and potentially helpful to be able to predict and imitate an algorithm, AlphaGo helps show how it&#x27;s even better to learn, as Fan Hui is attempting to do, what&#x27;s really going on behind the scenes.<p>Despite never having placed a Go stone in my life, I have particularly enjoyed Nick Sibicky&#x27;s Go lectures on YouTube. This one analyzing some of AlphaGo&#x27;s tendencies from its recent online matches is particularly interesting: <a href="https:&#x2F;&#x2F;youtu.be&#x2F;v8Eh41m7gVA" rel="nofollow">https:&#x2F;&#x2F;youtu.be&#x2F;v8Eh41m7gVA</a>
nerdponx超过 8 年前
This is a lot like constructing a partial dependence plot [1][2] for a very, very sophisticated predictive model.<p>1: <a href="http:&#x2F;&#x2F;scikit-learn.org&#x2F;stable&#x2F;auto_examples&#x2F;ensemble&#x2F;plot_partial_dependence.html" rel="nofollow">http:&#x2F;&#x2F;scikit-learn.org&#x2F;stable&#x2F;auto_examples&#x2F;ensemble&#x2F;plot_p...</a> 2: <a href="https:&#x2F;&#x2F;cran.r-project.org&#x2F;web&#x2F;packages&#x2F;datarobot&#x2F;vignettes&#x2F;PartialDependence.html" rel="nofollow">https:&#x2F;&#x2F;cran.r-project.org&#x2F;web&#x2F;packages&#x2F;datarobot&#x2F;vignettes&#x2F;...</a>
gort超过 8 年前
Link didn&#x27;t work for me, but Google brings up this:<p><a href="https:&#x2F;&#x2F;qz.com&#x2F;897498&#x2F;by-sparring-with-alphago-researchers-are-learning-how-an-algorithm-thinks&#x2F;" rel="nofollow">https:&#x2F;&#x2F;qz.com&#x2F;897498&#x2F;by-sparring-with-alphago-researchers-a...</a>