Enhancing the Decoding Rates of BATS Codes by Learning With Guided Information

Abstract

BATched Sparse codes (BATS codes) are a class of random linear network code designed for wireless multi-hop networks with packet loss. The encoder of a BATS code generates batches where each batch contains a number of coded packets. As the outer code is a matrix generalization of the fountain code, the ordinary batch construction scheme relies on a degree distribution with a random packet sampling scheme. In practical applications, we want a batch construction scheme which achieves a high decoding rate at the destination. A natural question to ask is: Is there any batch construction scheme which achieves a higher decoding rate than the ordinary one? We give an affirmative answer to this question by formulating the batch construction scheme as a multi-armed bandit problem and solving it with a deep reinforcement learning method with the degree distribution as a guiding prior. The BATS code generated by our proposed method achieves a higher decoding rate with improved efficiency compared with the ordinary batch construction scheme.