Quantum Assisted Scheduling Algorithm for Federated Learning in Distributed Networks
Quantum Assisted Scheduling Algorithm for Federated Learning in Distributed Networks
复制标题
DOI：
10.1109/icccn58024.2023.10230094
复制
发表时间：
2023-07
期刊：
影响因子：
--
通讯作者：
Xinliang Wei;Lei Fan;Yuanxiong Guo;Yanmin Gong;Zhu Han;Yu Wang
中科院分区：
文献类型：
--
作者：
Xinliang Wei;Lei Fan;Yuanxiong Guo;Yanmin Gong;Zhu Han;Yu Wang
在分布式网络中使用多个模型的联合学习（FL）的调度问题具有挑战性，因为它涉及NP-HARD混合组合非线性编程。此外，它需要多个FL模型之间的最佳参与者选择和学习率确定，以避免高培训成本和资源竞争。为了克服那些chal-lenges，在文献中，弯曲者的分解算法（BD）可以解决混合整数问题，但是，它仍然具有有限的可扩展性。为了解决这个问题，在本文中，我们介绍了混合量子式弯曲器的分解（HQCBD）算法，该算法结合了量子和经典计算的力量，以解决多模型FL中的联合参与者选择和学习计划问题。 HQCBD将优化问题分解为具有二进制变量和带有连续变量的小子问题的主问题。这种协作最大程度地提高了量子和经典计算的潜力，并优化了复杂的关节优化问题。商业D波量子退火机上的仿真证明了该方法的有效性和鲁棒性，即使在小规模下，迭代率提高了18％的迭代提高，计算时间比BD算法改善了81％。
The scheduling problem for federated learning (FL) with multiple models in a distributed network is challenging, as it involves NP-hard mixed-integer nonlinear programming. Moreover, it requires optimal participant selection and learning rate determination among multiple FL models to avoid high training costs and resource competition. To overcome those chal-lenges, in literature the Benders' decomposition algorithm (BD) can deal with mixed integer problems, however, it still suffers from limited scalability. To address this issue, in this paper, we present the Hybrid Quantum-Classical Benders' Decomposition (HQCBD) algorithm, which combines the power of quantum and classical computing to solve the joint participant selection and learning scheduling problem in multi-model FL. HQCBD decomposes the optimization problem into a master problem with binary variables and small subproblems with continuous variables. This collaboration maximizes the potential of both quantum and classical computing, and optimizes the complex joint optimization problem. Simulation on the commercial D-Wave quantum annealing machine demonstrates the effectiveness and robustness of the proposed method, with up to 18% improvement of iterations and 81% improvement of computation time over BD algorithm on classical CPUs even at small scales.