Mutagenesis: Retropostspectmortivem

gcewing on 2011/10/05 05:17

I didn't have much time to spend on my entry this time, so making it into the top half of the rankings was better than I expected. Thanks to all those who rated.

It seems that most people didn't find it to be all that much fun, and I can understand that. The whole project was highly experimental; I didn't really know how the idea would work out when I started.

One thing I think I've learned is that while L-systems are good for designing organisms by hand, they're not so good for random mutation experiments. One reason is that they have quite a lot of structure to them that must be preserved for the result to be well-formed, which makes the mutation algorithm rather messy and complicated.

Another problem is that most mutations tend to either do something that's not very interesting or render the organism unviable. I ended up putting quite a few hacks into the mutation algorithm in an effort to steer it towards interesting mutations, but I wasn't very successful.

I'm not sure how much of this is due to the choice of L-systems for representing the genes, the scheme used for interpreting them, or the nature of genetics in general. It's interesting to note that the other entry along similar lines, Lab Lab Bunny Lab, seems to exhibit a similar problem. LLBL appears to use a less structured representation for the genes, although the interpretation is still something turtle-graphics-like.

If I were to attempt another project along these lines, I would probably try to model more directly the process by which organisms grow and develop under the direction of genes. I would also try to make
the gene encoding dense and as devoid as possible of any necessary structure, so that just about any mutation would produce something viable.

On the other hand, it may be that we're imitating real life too closely. Nature has to roll the dice billions of times before hitting on an improvement, and doesn't care how many creatures it kills in the process. Devising a game mechanic that mimics evolution but produces interesting results after only a few tries appears to be a significant challenge.

(log in to comment)

Comments

Hugoagogo on 2011/10/05 06:45:

Isnt that more or less representative of standard evolution, many microscopic changes over thousands of years, and if you get unlucky and are an unfit mutation you die

gcewing on 2011/10/05 06:49:

Yes, that's my point. We're trying to compress a long and tedious process into a short space of time and make it fun. It's not surprising that we're having a few difficulties!

Cosmologicon on 2011/10/05 15:26:

Okay, but none of the games in this competition (that I can think of) were really analogous to natural selection, in that the fitness was not determined by the environment. They were more analogous to artificial selection, aka breeding. This process is generally much faster. It doesn't take thousands of generations to breed desirable traits. I think 5-10 generations would be enough to see some results.

None of the games had a super responsive interface, though, so even 5-10 generations was not very easy to get through. A more responsive interface would do the trick, I think. There's an experimental game called Darwin Hill that I think does a good job with this. It only takes them about 2 minutes to breed two distinct populations of creatures in this video:

http://www.youtube.com/watch?v=8enaDpxcWvI

Hugoagogo on 2011/10/07 07:00:

It would have been fun to see a game where you have to mutate a species by introducing hazards into the enviroment