How to build a chess engine

This is pretty tremendous: https://codegolf.stackexchange.com/questions/19397/smallest-...

Chess computing is one area where there’s a lot of attention to hard details as well as the algorithms, lots of gems like this in the genre

Yes it does.

I think chess engines are an amazing thing to study. So much fundamental computer science is covered, from extremely low level CPU instructions to high level data structures and algorithms. There's no surprise to me that many of the greats have turned their eye to representing and solving chess, from Turing to Knuth. I strongly suspect you could teach an entire CS degree never stepping outside of chess.

Even now there are so many rich directions you could go in. For example, many people are working on explainability in machine learning, but I think in many applications you want to go even further and actually optimise a model to _teach humans_ not just explain its predictions. This is especially relevant to engines when making strategic moves rather than calculating short term tactics.

I also think engines as tools for match preparation could be much better. What if you could more easily set an engine to find forced draws in certain lines, or tailor an engine to find tricky lines that a particular opponent is less likely to follow correctly, or find and optimise for certain endgame structures an opponent is weak at? Doing this in ChessBase (a truly awful but necessary piece of software) is quite painful.

Chess programming is fascinating but it seems for me that it has somewhat stalled in creativity in terms of classical (non DNN) engines. They all seem to use improved versions of min-max/alpha-beta. Which results in computer players that are extremely powerful but also quite dull. This is because they assume that the opponent will play perfect game and will do it at the full strength of the engine.

Basically there is no point playing against the engines at full strength as a human, since it will beat almost any human beyond grandmaster level. And that’s for weaker engines – the medium tier ones will regularly beat grandmasters and the top ones around the stockfish tier will probably not loose 1 game in 100 plays.

That considered I’ve wondered what’s possible dropping the assumptions about perfect play of the opponent. Some probabilistic models of opponent or different way of rating whole tree of moves in the context, instead of position only. For example considering the alternative moves and how difficult it would be to play perfect game for the opponent, it might be better to play combination with a lot of traps with possibility of loosing some position if the opponent plays perfectly. Current engines wouldn’t play such kind of move because they’d assume (via min-max/alpha-beta) that opponent will have perfect answer. Of course, if the opponent is an engine it will, but against humans it would make things much more interesting.

You might find https://www.youtube.com/watch?v=DpXy041BIlA interesting. He creates a tournament of "imperfect" engines.