ZGS-Based Event-Driven Algorithms for Bayesian Optimization in Fully Distributed Multi-Agent Systems

Pengfei Ren, Cheng-zhong Xu

Primary Area: optimization

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: distributed machine learning, Bayesian optimization, multi-agent systems, zero-gradient-sum optimization, event-driven mechanism

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Bayesian optimization (BO) is a well-established framework for globally optimizing expensive-to-evaluate black-box functions with impressive efficiency. Although numerous BO algorithms have been developed for the centralized machine learning setting and some recent works have extended BO to the tree-structured federated learning, no previous studies have investigated BO within a fully distributed multi-agent system (MAS) in the field of distributed learning (DL). Addressing this gap, we introduce and investigate a novel paradigm, Distributed Bayesian Optimization (DBO), in which agents cooperatively optimize the same costly-to-evaluate black-box objectives. An innovative generalized algorithm, Zero-Gradient-Sum-Based Event-Driven Distributed Lower Confidence Bound (ZGS-ED-DLCB), is proposed to overcome the significant challenges of DBO and DL: We (a) adopt a surrogate model based on random Fourier features as an approximate alternative to a typical Gaussian process to enable the exchange of local knowledge between neighboring agents, and (b) employ the event-driven mechanism to enhance communication efficiency in MASs. Moreover, we propose a novel generalized fully distributed convergence theorem, which represents a substantial theoretical and practical breakthrough wrt the ZGS-based DL. The performance of our proposed algorithm has been rigorously evaluated through theoretical analysis and extensive experiments, demonstrating substantial advantages over the state-of-the-art baselines.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9310