Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a transformer-based approach within the PatternBoost framework to generate Hadamard matrices, focusing on the Goethals-Seidel type to leverage fast Fourier-based scoring. The method outperforms local search from random initializations, particularly for orders between 100 and 250, producing numerous inequivalent Hadamard matrices, including one of order 244. The study also reveals the transformer's ability to exploit hidden symmetries within the combinatorial search space.
Transformers can discover hidden symmetries in combinatorial search spaces, enabling the generation of Hadamard matrices that elude traditional local search methods.
We present a new method for constructing Hadamard matrices that combines transformer neural networks with local search in the PatternBoost framework. Our approach is designed for extremely sparse combinatorial search problems and is particularly effective for Hadamard matrices of Goethals--Seidel type, where Fourier methods permit fast scoring and optimisation. For orders between $100$ and $250$, it produces large numbers of inequivalent Hadamard matrices, and in harder cases it succeeds where local search from random initialisation fails. The largest example found by our method has order $244$. In addition to these new constructions, our experiments reveal that the transformer can discover and exploit useful hidden symmetry in the search space.