When algorithms surprise us | Hacker News Reader

When algorithms surprise us

arxiv.org

112 pointsimartin2k18 comments

I think it was almost 20 years ago when I expectantly started the robot algorithm that had slowly been evolving for weeks.

I had already realized that I had to adjust the fitness function to give points for killing, not just surviving, since the first attempt had created robots that instead of skillfully picking off the opponents one by one - as I had envisioned - took speed for the nearest corner and then just sat there waiting to be killed by one of the other traditionally programmed robots.

I realized that the corner strategy made sense as they got the most protection there, but it was a bit unsatisfactory to watch...

In the second try the robots scurried to their usual corners, waiting to be killed, but this time firing randomly, occasionally killing one of their opponents by luck before they got hit.

The evolutionary algorithm sort of hit a plateau there, and besides I kind of wanted to use my computer for other things rather than spending days and days simulating robot fights.

But I learned two things:

* Evolution most likely is a thing.

* It's really really hard to write fitness functions...

So you were working on a lab proof of S.L.A. Marshall's Men Against Fire? ;-) https://en.wikipedia.org/wiki/S.L.A._Marshall

It's important to make sure that what these algorithms actually produce is surprising to subject matter experts. I can recall, but can not find at the moment, a tumblr post where the blogger was surprised that a computer program which optimized a physical structure to minimize material while maximizing strength ended up looking organic.

This didn't surprise me at all. Organic structures evolved over millions or billions of years and probably are nearly optimal at accomplishing a particular task. I'd be surprised if the optimization software didn't produce something that looked organic.

It's not like the optimization software was better than actual organic structures either, which are neither isotropic (in organic structures the strength varies depending on the direction) or homogeneous (in organic structures the strength varies depending on the location) as the software had assumed.

This article is actually providing select anecdotes from a more exhaustive paper: https://arxiv.org/pdf/1803.03453.pdf

Ah, good catch. It was discussed here: https://news.ycombinator.com/item?id=16600701. But maybe a second discussion isn't so bad.

We'll change the URL to that from http://aiweirdness.com/post/172894792687/when-algorithms-sur....

The title is still the same as the original link, but I wanted to comment on the use of the term "algorithm" in this sense. I'm wondering if that's the proper term for the output of a machine learning algorithm, and if we shouldn't be calling it "learned behavior" instead. I just feel it disparages all the careful work of library developers and the like and cheapens an important field of study.

I tend to believe life is universal. It may arise in a Turing machine that executes all possible programs one line at a time in parallel. It may arise in the subatomic realm... assuming sub-atomic particles exhibit sufficiently diverse complexity between 10^-35 and 10^-17 meters (a space unknown to us). It may arise between galaxies in the universe etc. We should view life more broadly.

But the one thing that binds everything in the universe is that we must eat. And this constant depletion of the entropic energy in our universe comands us to see as much as possible before the light burns out, lest the universe never will never know we existed at all.

Then I think that you would enjoy "Last and First Men and Star Maker" by Olaf Stapeldon.

Available quite reasonably at https://www.amazon.com/Last-First-Men-Star-Maker/dp/04862196....

I'm also reminded of https://blog.openai.com/faulty-reward-functions/ (which I saw on HN a while ago), where an AI unexpected learned to

> turn in a large circle and repeatedly knock over three targets, timing its movement so as to always knock over the targets just as they repopulate. Despite repeatedly catching on fire, crashing into other boats, and going the wrong way on the track, our agent manages to achieve a higher score using this strategy than is possible by completing the course in the normal way.

Would be interesting to provide videos of the evolved robots rather than just the anecdotes.

The paper he draws the examples from has links to videos of most of the visually interesting examples.

Very interesting. 'Learning to Play Dumb on the Test' was the most surprising one to me, followed by the alien-turned-car that lead to 'novelty search' algorithms.

> Sometimes I think the surest sign that we’re not living in a computer simulation is that if we were, some microbe would have learned to exploit its flaws.

Someone needs to tell the author about quantum mechanics. It's not an unreasonable hypothesis to explain quantum effects as numerical errors in the Matrix.

Thought the same! What if life is exploiting such a flaw - a flaw that led to the existing matter out of which our universe is built.

Also the speed of light is dictated by how often the simulation "ticks".

It doesn’t make sense for quantum effects to represent errors as much as it would for them to be “shortcuts” to save on power. This would be especially true if spacetime itself turns out to be quantized rather than continuous. Either way though, the probabilistic nature of QM on its own could represent a “fudge” factor I guess?

Truthfully simulation theory is uncompelling as a real theory, but very entertaining to contemplate.

> It's not an unreasonable hypothesis to explain quantum effects as numerical errors in the Matrix.

Have you studied quantum mechanics? Numerical errors in a simulation do not produce similar effects.

For example, consider Shor's algorithm. If QM is just classical mechanics + error, why can't we use a bad physics simulation to factor numbers? Where would the work needed to do the factoring even be happening?

Reply on news.ycombinator.com