%2BCelo%5B%5Cs%5CS%5D*%3F%5Cb(%5Cd%2B)(%3F%3A%5Cs%7C%26nbsp%3B)%2BIntegral%5Cb&replace=Rank%20%23%241&label=sp-cc.de&color=blue&cacheSeconds=86400){kind=icon}
Integral is a top chess engine developed in C++. It is a personal project developed with the goal of meshing my admiration for chess with programming.
Integral implements the widely adopted negamax search approach with alpha-beta pruning, and alongside it the various search heuristics that it enables. It utilizes the Lazy SMP approach for multi-threaded search, and has been proven to scale very well at higher thread counts compared to other alpha-beta chess engines.
Integral utilizes an efficiently updatable neural network (NNUE) for its evaluation function.
Integral's neural network is a horizontally mirrored perspective network, containing 16 factorized king input buckets with threat inputs, pawn-pawn pair inputs, and fifty-move rule inputs, an L1 of 768 neurons, an L2 of 16 neurons, an L3 of 32 neurons, and 8 output buckets.
[768x16hm (Factorized) + 1x11 (HMC) + 59808hm (TI) + 4560hm (PP) -> 768]x2 -> (16 -> 32 -> 1)1x8
This neural network is trained on hundreds of millions of self-play games. Most self-play games starts with 3-4 randomly selected moves off a randomly selected opening from the UHO_Lichess_4852_v1 book. Additionally, a majority of the data uses 5-man Syzygy endgame tablebases to guide the data generation search.
The first iteration of Integral's neural network was trained on data from version 4, which had a powerful hand-crafted evaluation (HCE). All early-network data has been thrown out and a majority of Integral's data is generated using a search of 20k soft-nodes per-side. The datasets Integral trains on is publicy available through my Hugging Face repository. Networks are trained using the Bullet trainer.
Note
Integral should be compiled with GCC >= v13 or Clang >= v10
Integral provides a Makefile for compilation, but you can use CMake as well.
To compile Integral, enter the following commands in a terminal:
git clone https://github.com/aronpetko/integral
cd integral
make [native | vnni512 | avx512 | avx2_bmi2 | avx2 | sse41_popcnt]
Clang is used when available, and builds use profile-guided optimization by default (disable with PGO=OFF).