signBEER: Decentralized Optimization with Improved Communication by Signs

Abstract

This paper presents signBEER, an algorithm designed to enhance communication efficiency within the BEER algorithm framework (Zhao et al., NeurIPS 2022). In contrast to the preceding BEER algorithm, signBEER utilizes communication vectors composed exclusively of signs. Our findings reveal that signBEER outperforms the original BEER algorithm, resulting in enhancements in communication efficiency. The realized cost savings with signBEER are higher than those achieved with the traditional BEER algorithm. Furthermore, signBEER exhibits greater resilience against communication errors compared to BEER, attributed to its reliance on signs. Our proposed algorithm, signBEER, shows particular promise in scenarios with constrained bandwidth, privacy enhancement requirements, or error-prone communication, such as in IoT environments.