Public:Digital Evolution
From ZevilsWiki
Life in the Machine: Experimental Evolution with Digital Organisms; talk by Charles Ofria. Notes by Matthewg 14:03, 30 October 2006 (PST)
Contents |
Thought process
- Artificial intelligence. Really big hurdles, field is having difficulty living up to promise. Thought abaout what I would do differently.
- Evolution. Only one real instance of intelligence in the natural world, so started studying evolutionary biology. Tried to understand the rise of intelligence. But evobio had lots of interesting open questions, so...
- Went back to computer science to try to understand it.
Evolution is hard to study
Living systems are messy.
- They reproduce slowly. Experiments with evolving populations require waiting a long time
- They interact at many different levels
- They're fragile
So, let's have evolution running inside a computer.
Digital Organisms
- Started out in the Core War game. Most successful strategy is to have self-replicating programs so that if a copy somewhere in memory gets destroyed, there's another elsewhere. No evolution because there's no source of variation.
- Tom Ray's 1991 Tierra was like Core War but with mutations. We saw evolution. There was an 80-line version of the program, mutation created a 79-line version which outperformed the 80, eventually got down to 65 lines.
- Suddenly, a 45-line parasite emerged.
- There was an arms race, cyclic population behavior...
- Then, a 78-line program emerged and wiped everyone else out. It laid traps for the parasites.
Avida
I made this in 199X. A research program for evolutionary processes. Good data collection and analysis tools.
- Organisms use energy in the form of CPU cycles. In simple environments, all organisms receive fixed energy.
- In complex environments, organisms can compute math functions to "metabolize a resource" and gain energy.
- Gentle slopes and sharp jumps. Gentle slopes are refining what they can already do (e.g. "one line shrinking"). Sharp jumps correspond with new tasks that give them access to a new resource. Declines are rare.
Competition Experiment
Two organisms, one (yellow) evolved in a normal environment, one (blue) evolved in a high-mutation-rate environment. So, blue is "worse" than yellow but more robust. We pit them against each other in yellow's environment; yellow overwhelms blue. In blue's environment, blue wins. (This is without crossover.)
I wanted to run tests without new things evolving during our tests to taint results, so I set it up so that whenever there was a beneficial mutation we killed the organism. But things become more fit anyway. WTF? Turns out that my test environment (for testing whether the mutation was beneficial) was too clean and was always using the same numbers, and organisms evolved to recognize those numbers and "play dead" when they saw the test data. It's frustrating when you're outsmarted by 200 lines of code. Okay, I'll make the test population look just like the real one, same distribution of chaos... That worked for a while, but eventually they made it so that they would only perform a task 50% of the time.
Applying new systems to old questions
Darwin recognized three major challenges to his theory:
- The habitable age of the Earth. Darwin felt that the Earth wasn't old enough for his theory to work (because they were wrong about the age of Earth since they thought that the sun was a big burning pile of coal.)
- Why aren't children just a "blend" of their parents? (They didn't know about crossover.)
- Origin of complex features like the eye. (Still to some extent an open question, although not with the eye.)
Studying the origin of evolutionary complexity
We did 100 experiments, 50 in a "reward-all" environment with 9 different tasks, 50 in a control where only very complex tasks were rewarded. At the end, 23/50 reward-all acquired the complex feature. None of the controls acquired it. But how was it (the ability to perform bitwise equality comparison) acquired?
(Hm, I'd like to hear the details of how they're measuring the fitness — reproductive speed — of an organism...)
In each of the 23 the acquired it, it happened a different way. In 20 of them, when they acquired it, it was through giving up some lesser task.
