Hilarious. I love generational learning like this (though I feel bad for the entities).

My friend and I one night spent about 2 hours in front of my laptop live-commentating the BreveCreatures screensaver, in which procedurally generated polygonal creatures attempt to walk, and pass on patterns and configurations to the next generation.

"Looks like we've got another flopper here. That might have flown a couple generations ago but you know, we've been seeing some really effective writhing lately."

"I totally agree. The bipedal taildraggers were a force to be reckoned with in season 56, but — oh my god! Look at that one go!"

"Almost got as far as that three-legged galloper we saw just a few minutes back. This is a heck of a race, my friend!"

For hours. We were drunk.

I would invest a bit to encourage you and your friend to put this on youtube. : )

I bring this up because these things are really entertaining, and with a bit of witty commentary and nice graphics, I would probably end up spending hours watching these. Like all stories, you have to have an 'arc', themes, etc, but these come later.

In college, I must have spent thousands of hours watching let's play for SimCity and Minecraft. My friends were always amused how I enjoyed watching the gameplays more than actually playing those games.

On the other hand, you don't need me at all.

Edited for clarity.

I would love to see the youtube for that too :)

Are you sure you were not also stoned

I've noticed that those who bunch together reproduce very well but the planks sometimes form death traps killing the entire group. Some creatures are left over but they seem to avoid each other; sometimes one creature is chasing the other and they don't seem to reproduce very much.

Bunching and spreading out are two different behaviors which both have their pros and cons. I wonder if the neural net is complex enough to enable them to switch between these behaviors depending on the age and ability to reproduce.

I wondered the same thing; I suppose it depends on how frequent those kinds of mass extinction events are. You might end up with two species:

1. The kinds that clump and are subject to rare mass extinctions, but get along well enough otherwise.

2. The kinds that are really good at avoiding converging obstacles by avoiding clumps.

Edit: I just realized that the clumps are kind of a communication network and eventually you might be able to get a kind of clump that either is able to quickly evade converging obstacles or be self-sacrificing to save a part of the whole.

I guess (1) or (2) would depend on the culture that gets evolved :)

The network would be too small to develop this complex behavior.

I had to balance between having it run well in js and the computation complexity of each creature.

Is the ability to reproduce an input value to the network?

pretty funny how sometimes the agents move away from the planks like in a "The Prometheus School of Running Away From Things" fashion

I liked how some of them would "impale" themselves on the end of a plank, like they couldn't see it if it was pointed along their direction of travel.

It's probably because they have eight "sensing rays" and plank may be located between them.

Yeah, I can make them have 16 but it slows things down esp on mobile phones.

Evolutionary algorithms may work ok here, but I have not seen any high-performance ML tasks done with evolutionary training of neural networks in a long time. It's all backprop and, increasingly, reinforcement learning these days.

Have you ever looked at https://github.com/karpathy/reinforcejs ? Karpathy already has demos similar to yours which you could build on. The trained agents might be better.

Yeah reinforcement learning and policy gradients can possibly attack the same problem.

Although for a simple task like this demo, I stand by simple conventional neuroevolution as they would work best and are easy to train as the nets are quite small.

> Although for a simple task like this demo, I stand by simple conventional neuroevolution as they would work best and are easy to train as the nets are quite small.

They may be easy to train but I'm not sure about 'work best'. Watching your final trained agents, I see a lot of dumb obvious mistakes where the agent runs right into a line; and I see little higher-level strategic planning like trying to maximize free space around one or run away from a incoming line.

I’m not sure why, but the page is completely unreadable: http://i.imgur.com/YgP34SE.png

And this is with Firefox, which usually renders fonts a bit fatter than competing browsers.

For some reason they are using this style which is responsible for that:

  font-weight: 200;

FYI, the common range for font-weight style is 100-900, and the normal font weight is 400.

Looks fine on my Mac: http://i.imgur.com/wxZAbgk.png

Well, you have 3.5 times as many pixels, of course fine fonts are readable there.

Hi Kuschku

Can you let me know the resolution of your device? I'm trying to play around with the CSS for blog.otoro.net, and trying to find a good balance for different resolutions.

Thanks

Seems to be fixed now; looks better on my FF

Thanks for the feedback, I'll need to change around the CSS later.