Optimized Kernels in MobileNet Architectures for Computer Chess

This work investigates the use of various Residual and MobileNet architectures within the Leela Chess Zero (Lc0) engine. In particular, it examines convolutional layers using specific kernels inspired by chess piece movement patterns. The most effective configuration, combining knight, rook, and bishop filters, is further optimized at a low implementation level to speed up network inference during move search. All network variants are trained on the recent T80 dataset, consisting of self-played games generated by Lc0. Their relative performance is then evaluated through a series of tournaments conducted under different time controls, including bullet, blitz, and rapid formats.