HackerTube

Ken Stanley: Discovery Without Objectives

fzambetta751 HN recommenders · 2016–2018
6K views · talk · evergreen · Uploaded 2012-09-04

A 92-minute RMIT University seminar in which Ken Stanley argues that setting an objective can block achieving it. He presents objective-less, novelty-driven search as an alternative approach to innovation and discovery, drawing on evolutionary computation.

Cited as a source for the idea that open-ended search can beat objective-driven optimization, though one commenter found parts of it fluffy.

Machine learning & deep learningScientific method & science communicationEvolution & genetics

Watch on YouTube

3 comments from Hacker News

@greggman2016 · in “Why Greatness Cannot Be Planned [video]”

Interesting talk. I really couldn't tell how much of it really meant something and how much of it was fluff like "power poses". In any case he's been giving similar talks for a while. [VIDEO]

CriticalView on HN
@truculation2018 · in “You Can Always Get What You Want, But Not What You Need (2016) [video]”

Fascinating. Instead of banning money or fixing prices (did they try that in Russia?) we retain free markets but now with a new layer of government intervention. Not intervention by social spending or regulation but by buying and selling according to the dictats of an AI. And the utility function of this AI is set according to what exactly? My guess is it would have to be according to our values, not to specific goals: [VIDEO]

CitesView on HN
@mkl2018 · in “Generative Art Finds Its Prodigy”

I don't think this algorithm is very common. In particular, the colour palette can usually be chosen last , and a lot of generative art (at least all that shown in the article) will not need multiple resolution passes. I think a more common discovery/refinement process is to go from more random to less random, or random in particular kinds of ways. For example, you can specify the values or distributions (or at least ranges of values) for particular parameters based on things you think will look good or which should end up similar in ways you like to things you've found already. Generative art is a search in an extremely high dimensional space. The more you go into it aiming for something in particular, the more nailed down it is and the closer to traditional art (and the harder to pull off well IMHO, think how much difference slightly changing the position of one of the elements in these images would make, versus slightly changing the position of someone's eye). If you go looking for "new and interesting" or "looks good" without really aiming for anything in particular, it can be a lot easier to find good things. I've seen a couple of really interesting talks recently about the surprising successfulness of objective-less search by Ken Stanley: [LINK] (40 min.) [VIDEO] (90 min.) He started off looking at human-directed machine-evolved generative art, and ended up with a far more general discovery.

RecommendsView on HN