Learning To Branch-And-Bound For Header-Free Communications

In this paper, we present a learning-based approach for solving shuffled linear systems in header-free communication, thereby supporting low-latency communication. The resulting shuffled linear regression problem aims at solving a linear system with unknown permuted measurements, which is known NP-hard. Although global optimization algorithms such as branch-and-hound algorithm can obtain the globally optimal solution, its computational complexity is exponential and can not he applied in real systems. To alleviate the computation burdens, we propose to learn the pruning policy in the branch-and-bound algorithm. Specifically, we first formulate the branch-and-bound algorithm for solving shuffled linear systems as a sequential decision problem, followed by learning the optimal pruning policy. It is shown that the expected computational complexity of the proposed algorithm is polynomial with careful features design. Numerical experiments demonstrate that the proposed method both achieves near-optimal performance and significant speedups to the branch-and-bound algorithm.