Chess Engine

By far the biggest and most challenging personal project I’ve worked on. Creating a chess engine from scratch has been one of my biggest personal goals, so my motivation was plenty. Since chess is a game that I actually play often, I enjoyed injecting some of my own strategy and knowledge into my program.

It began with my choice to pick up a lower level language before beginning this project. Although I enjoyed programming in Python and JavaScript, I knew that I would have trouble getting good performance with their many levels of abstraction. I decided to get muddy with C++, because although the learning curve is high, I was attracted to the large standard library and was used to C-style syntax. It took me many days before I got a good enough grasp on the language.

I spent a while on the design decisions (move/board representations, class hierarchies, etc.). It was the first time I felt like I was making something big. I decided to incorporate the object-oriented nature of C++ into my design decisions, while making sure that no performance would be hindered. After weeks of not super fun work, I had the skeleton of a chess engine, move generation, complete. After extensive testing, I was sure that I had implemented all the complex rules of chess correctly. The fun part could begin.

I had implemented the minimax algorithm half a dozen times by now, so I was excited to experiment with ways to make it faster. I learned about the iterative deepening framework, and how it can be used with transposition tables and move ordering to make alpha-beta pruning extremely effective. I learned about killer moves, PV/null-window search, and search extensions. I found myself going down rabbit-holes on chessprogramming.org for hours. With a relatively simple static evaluation function, I implemented many of these optimizations, and found myself quite the formidable opponent.

I collected thousands of chess positions to test the engine with, so I could tune parameters and change methods in the search algorithm. I'm working on a sparring-program to have two different versions of my engine play against each other in a variety of positions. The current version is stronger than I ever expected I could accomplish.

UPDATE: Interested in the modern machine learning approaches to chess engines that are so common today, I went on the journey of creating my own neural network evaluation function. Since my engine is inspired by stockfish, I dove into their documentation. I read all about their very unique approach to a neural network evaluation function, NNUE. They utilize a very unique network shape and board vectorization that allows for very efficient inference. I decided it would be fun to try and do it on my own.

I collected a large amount of position/evaluation pairs from across the internet, threw them together in python with pandas, and trained a neural network model using pytorch. I then extracted the weights, and went about the process of integer quantization. I coded my own implementation of the trained network in C++ using intrinsic vectorized CPU operations of the quantized network. You can view more about the process HERE.

I don’t have anything that you can play online right now, but feel free to check out the source code, and see what features I am currently working on. I wrote a simple chess GUI so you can play against my latest engine.