A few years ago I decided that to fully understand the dynamics of genetic programming, I would have to implement my own framework. Beagle had proved incredibly useful, but I required full control of the entire software design, down to minute details, so I had to write my own.
Named after a dog we had when I was a kid, and after a certain pioneer in the field of evolution, I called my new framework Charlie (if you were wondering why this blog was called CharlieAI, you now have your answer!)
I will spare you the implementation details and instead list the feature set I ended up with:
- Tree-based genomes
- Cross-over, mutation and re-selection operators
- Node and sub-tree mutation
- Tournament-based selection
- Ephemeral random constants
- Population statistics
- Lisp-style and graph in .gif file output (Graphviz)
- Multi-threaded evaluation
- Single and co-evolution
- Hall-of-fame archives
This was all built over six months or so, but I only had a few hours of time to devote every week so the whole thing could easily have been done in something like two weeks of full-time work. I must admit, though, that it was nice to have ample thinking time between implementation steps, to ensure it all fitted well together and performed efficiently.
Next I will cover the results I got from running symbolic regression experiments with this framework. In the meantime, if you would like any more details on any of the above features or any aspects of the software, do let me know and I will attempt to cover it in a future post.
No comments:
Post a Comment