I am curious why the author chose a genetic algorithm rather than standard backprop to distill the evil. Logistic regression seems like a pretty reasonable choice and it’ll be a lot faster than a genetic algorithm. Add an L1 penalty for sparsity.

In the past I’ve tried distilling not just the eval function but the outcome of the search (something like depth 20) in a neural net. It kind of works, but not very well until the net is pretty big. Deepmind did something similar.

I especially enjoyed the link to The Bitter Lesson by Rich Sutton, which I hadn't read before. Now I wonder what "discoveries" have been built into today's AI models and how they might come to be detrimental.

http://www.incompleteideas.net/IncIdeas/BitterLesson.html

All this I guess comes after laying the ground work, like implementing a bitboard representation or something else, and implementing the logic for being able to tell, whether a move is valid or invalid. That in itself is already a lot of work. iiuc the idea here is "merely" writing the engine, and take everything else as a given.

I built a list of simple games in pygame with code and solver generated by AI.

Have implemented 2048, Minesweeper, Sudoku and chess. First three are single player games have made good progress.

I still don't understand UCI interface and have not thought through chess solving. Hope will take another try this weekend

https://github.com/muonium-ai/simplegames

As part of my undergrad work, I used similar principles to the article (steady state genetic algorithms) to create a bot capable of playing Reversi where the fitness function was loosely defined as a set of "weights" on each square across the 8x8 board. These were used as part of the evaluation function in the classic minimax algorithm.

We trained over the course of 5-6 days, and the end generation was capable of beating an intermediate player but got completely destroyed by experts. It was a fun project!