Breakthrough in quantum optimization uses binary autoencoders for faster solutions

Researchers from Keio University have developed a new method to improve how complex optimization problems are solved. Their approach uses binary autoencoders to create better representations of solutions, making the search process faster and more efficient. The team, including Tetsuro Abe, Masashi Yamashita, and Shu Tanaka, focused on overcoming a key limitation in existing quantum optimization techniques.

The study introduces a framework called bAE+FMQA, which combines binary autoencoders with Factorisation Machines for Quantum Annealing (FMQA). Unlike traditional methods that rely on manually designed binary codes, this system learns a compact binary representation of problem solutions. This reduces the need for human input and improves how well the structure of the solution space is preserved.

The researchers tested their method on a traveling salesman problem, a classic optimization challenge. Their results showed that the learned binary codes could accurately reconstruct feasible solutions. The new approach also aligned tour distances more closely with Hamming distances in the latent space, reducing the number of invalid solutions generated during optimization.

By using the binary autoencoder, the team avoided the inefficiencies of manual encodings, particularly when dealing with non-binary structures like integer permutations. The learned representation improved search efficiency, offering clearer insights into how latent spaces can be designed for black-box optimization. The study demonstrated both higher approximation ratios and smoother neighborhoods when making small changes to the binary code.

The findings provide a practical way to enhance optimization performance on Ising machines and similar systems. The framework eliminates the need for handcrafted encodings while maintaining solution quality. This advancement could help solve computationally expensive combinatorial problems more effectively in the future.