A Visual Guide to Evolution Strategies | Hacker News Reader

A Visual Guide to Evolution Strategies

blog.otoro.net

260 pointshardmaru21 comments

“Wall clock comparison. Instead of looking at the raw number of states seen, one can argue that the most important metric to look at is the wall clock time: how long (in number of seconds) does it take to solve a given problem? This quantity ultimately dictates the achievable speed of iteration for a researcher. Since ES requires negligible communication between workers, we were able to solve one of the hardest MuJoCo tasks (a 3D humanoid) using 1,440 CPUs across 80 machines in only 10 minutes. As a comparison, in a typical setting 32 A3C workers on one machine would solve this task in about 10 hours.” https://blog.openai.com/evolution-strategies/

I would love to see this with a more complicated search space. The two examples are too easy to climb.

Would be really interesting to see how the different algorithms travel through a landscape where it is hard to find the optimum.

And to make it perfect: Put information about the performance below each visualization.

Near the bottom he does show the results of running it on 100-D Rastrigin (and mnist before that) with stats and plots. Or do you mean something more complex even than that?

Well, the point of the guide is the nice 2-D visualization. Which he cannot do for the 100-D problem of course. It would be nice to make the same guide but with a harder 2-D landscape.

This article is the first of a series of articles about the subject. The next article will be about solving more difficult problems with these methods. Stay tuned-

The next article in the series, Evolving Stable Strategies, is also up.

http://blog.otoro.net/2017/11/12/evolving-stable-strategies/

For me the biggest advantage is that ES often lends itself better to multi objective optimization. Discussed here: https://news.ycombinator.com/item?id=15584858

80s are so in right now, first neural networks, now genetic algorithms :P

what next... a discovery of the limitations of machine learning?

I’m scared by what the Japanese are doing with Prolog…

Actually the advances in logic programming are quite compelling; answer set programming is pretty cool!

I'm not sure why it's having such a small impact though - perhaps the shear number of low cost programmers available means that no one has to use it. Although I would have thought that high cost programmers equipped with tools like these could better justify their pay.

What are they doing?

Right, someone needs to go and check what learning classifier systems can do on modern hardware! :3

I’m expecting a renewed interest in Arrificial Life once the AI folks start to realize that the current crop of AI techniques is a nice bag of tricks but that we’re still falling short of what was promised.

Seems to be the general trend in technology in general. Just look at NodeJS and Javascript ecosystem reinventing 90s. Now they are at the stage of JSON and Electron. Just wait until they develop a XSLT equivalent for JSON.

Without Javascript the mathematical expressions are not displayed. This makes some parts of the text incomprehensible.

Yeah, but if you just casually don't use Javascript on the web, you're not worth catering to.

https://en.wikipedia.org/wiki/Progressive_enhancement

Yes, I know what progressive enhancement is, my point is that it's not worth the effort of starting from literally just html and CSS to cater for the people so unwilling to work with you that they don't even have the second most basic web tooling enabled.

And what kind of cost-benefit analysis is ever going to tell you that it's worth catering for the less than 0.5% of people who turn off JavaScript?

"As long as you're fine with the site completely failing because the browser is too old, or too new, or the user's bandwidth is too constrained, or the server hiccups, or a firewall's security policy blocks it, or a dependency goes sideways, or you accidentally drop a semicolon somewhere, then sure," says consultant and author Eric Meyer, "it's OK. What you build won't be a part of the web continuum, and it will be needlessly fragile, but that's a choice you can make."

http://www.creativebloq.com/features/is-it-okay-to-build-sit...

I like it :P

Reply on news.ycombinator.com