Bitwise Evolution / News: Recent posts

Thomas and I have been off school for this past week, and I've got a little bit done on this project. Today I sat down and wrote a really sketch ID system that just creates hashes of the combined genes. It works well enough, and allowed me to create some buggy code to color the creatures depending on their ID number.
There is a major problem with it, sadly. It appears that entire groups of creatures randomly change color. I don't know what this is all about. Makes me a little nervous that I had an entire concept wrong because I can't find any technical error.
For the past few weeks, this project has taken the back-burner for Thomas and I. Thomas, being a senior, has had to deal with all sorts of applications, and I have been bogged down with homework and projects for my classes. Now that all of that is pretty much done, we can start working on the project every day as we should've been doing these weeks past.

We finally got the CVS repository working to a satisfactory level of efficiency. Source code that is being worked on can be found in the brunswick/src area of the CVS. At certain critical changes in the code, we will update the main brunswick section's code. At the moment, the GUI is rather sparse, only containing the map and a slider to slow down the operating rate. Thomas will be working on that part. I'll be chugging away at the chromosome and the documentation for the project's process.
We got together on Saturday for most of the day, and came up with some really nice ideas. Now the creatures actually have genes that provide sight through references to map points. For example, a creature might have a gene whose code is "M010 | 0", which would give the appearance of whatever is directly above it in the matrix.
There is a bit of a problem with this though, because the neural network must change its size depending on how many different inputs there are. So if the creature mutates on a gene that references a point two squares away, the neural network has to expand to accommodate the addition.
Something very interesting that happened while we were watching the creatures was the ancestral memory that was developed. Because a parent's neural network is passed down to the child, the descendents of one creature will behave in very similar ways. What we found is that the creatures would learn where the food was, and just move on those squares in succession, allowing for enough time to regrow before they came back. They actually ended up becoming specialized for a certain area for the layout of the food. It was really amazing.
After witnessing this development, we decided to see how smart they could get. Instead of having the food regrow at a constant rate, the food regrows proportionally to how much it has been eaten in the last 25 cycles. So if a plant hasn't been eaten in the last 25 cycles, it regrows at a very good rate, but then if it is eaten too much in a short period of time, it doesn't regrow with the same veracity. What we saw with that was the training and then un-training of the networks. The creature would learn where the food was, but it would eat too much, and it would become isolated from other food. This would result in it just sitting in one spot, and dying off because it would be punished if it went in any direction.
It's some really amazing data we're seeing considering how little time we've had to work on it.
-ocrickard