I'm learning Nim and experimenting with evolutionary algorithms to generate logical functions.
As a test-case, I'm attempting to map from the binary address of an image byte (represented as 3 n-bit integers corresponding to the X, Y, and C coordinates of the byte) to the byte value itself. This is done by a randomly-initialized DAG of basic logic gates selected from {AND, NAND, OR, NOR, XOR, XNOR}.
Right now, the DAG is optimized by trial-and-error through random mutations which alter either the type or the inputs of a random gate each iteration. However, I'd like to implement a smarter search/evolution procedure soon.
- Pixie for image loading and saving.
- Bitty (vendored in as
src/bittyand lightly modified) for arbitrary-length bitvector data structure.
- Smarter input choices
- Smarter function choices
- Some flavor of "backprop"?
- Some flavor of global reward?
- Allow gates to output to any gate not in their parents — right now, graphs are arbitrarily restricted to a particular flavor of DAG instead of any DAG.