This repo was originally meant to be a simple implementation of Darwin-Godel Machines in GridWorld. As I learned more about that project, it seemed solving gridworld was too simple to warrant that approach, so I experimented with implementing different classical approaches instead. These include:
- A-star
- Tabular Q-learning (for small grids)
- Deep Q-learning
I heavily used LLMs to speed up development. The code is entirely written through a mixture of chatgpt and cursor agent prompting. As such, there is some repetition. But the A-star implementation is textbook.
The experience of working this way was pretty disorienting at times. I realize the main bottleneck here is my own working memory and code reading-comprehension speed, both of which seem fixed. I guess I'll have to look for ways to improve those.