Genetic Algorithm Fun

[dariusk]

So I was watching this genetic algorithm thingie I saw on Boing Boing. It's pretty cool. Then the Boing Boing comment thread pointed me to the original Reddit thread for the app. I saw an amusing comment there from ThisIsDave:

Are you familiar with Hod Lipson's work? He does a lot of the same things at Cornell.

Extraordinarily cool.

Edit

A few cool

Comments

[purple_dj]

There was a another real-world experiment where someone used a genetic algorithm to program FPGAs for a certain circuit. The rig was designed such that it could reprogram the FPGA and score the circuit without needing to disconnect/reconnect programmers or with any human intervention.

In the end, their run did produce the already known optimal design, but it also produced a few other more optimal solutions that took advantages of flaws in the individual FPGA chip in the socket. That is to say, the program was mostly nonsensical to anyone trying to read it, and wouldn't work on other FPGA chips of the same type.

Of course, I can't seem to find links or references to this anymore. You'll just have to take my word for it. (Screw you, Reading Rainbow!)

[dariusk]

There has to be a tie-in there to the old "beware of premature optimization" saying. Pretty amazing, though.

[slarti]

I, personally, hate that saying, because it easily leads to insistence that no, it's still too soon to optimize, to the point where engineers can never be made to actually work on it and sysadmins are instead expected to cough up increasingly crazily beefy hardware to support the poor performance characteristics of the software.

Uh, not that I'm bitter or anything, mind you.

[dariusk]

Oh yeah, it's one of the most abused old chestnuts in the history of programming. "I am lazy, let's not optimize."

[dominika_kretek]

I heard this anecdote also. I would love it if someone could find the reference.

More generally, exploits of the type in the original post are what have made me increasingly skeptical of the whole notion of simulation, in particular the simulation of biological processes and the brain. Unless you are simulating the entire universe to arbitrary precision (something that is arguably impossible to do within the universe), you have to divide the universe into two realms, what you're simulating and and what you think is irrelevant. (we call this "abstraction.") Genetic algorithms seem to be very good at making what you think is irrelevant relevant. I don't think that divide-by-zero error is telling us anything new about physics, but it's telling us a lot about the simulation. Probably if you had a good enough simulation, you could begin to learn interesting things about real-world physics, but you would always also see artifacts particular to the simulation

[dariusk]

Very well-said. This PDF, a sample from Salen and Zimmerman's "Rules of Play", begins with an anecdote about simulation and reality:

http://tinysubversions.com/zimmerman.pdf

As game developers, we're pretty tuned-in to the notion that a simulation is generally just a point of view.

[dirkcjelli]

pppst: this is a simulation generated inside your brain. Your senses have some granularity, they transmit information via electrical impulses, etc.

[dariusk]

Better thinkers than myself have referred to magic as the art of exploiting glitches in our reality-simulation.

[ab3nd]

I remember reading about this too. At one point they decided that one part of the design was obviously useless (one wire going in, none coming out), so they removed it. The circuit promptly stopped working, as it had evolved to rely on that wire as something other than a connection to another part (probably a parasitic capacitance that relied heavily on the specific IC being used).

[ab3nd]

Now we see the wages of replying before doing the research:

http://www.damninteresting.com/?p=870
http://en.wikipedia.org/wiki/Evolvable_hardware

It wasn't one connection, it was no connections at all. Still crucial to the circuit, though.

The first link also covers the evolved antennas, which look like they were designed by spiders on crack.

[dariusk]

Yeah, evolved antennas actually make sense, since we can pretty reliably simulate the E/M conditions involved.