The real-time operation of large-scale infrastructure networks requires scalable optimization capabilities. Decomposition schemes can help achieve scalability; classical decomposition approaches such as the alternating direction method of multipliers (ADMM) and distributed Newtons schemes, however, often either suffer from slow convergence or might require high degrees of communication. In this work, we present new primal decomposition schemes for solving large-scale, strongly convex QPs. These approaches have global convergence guarantees and require limited communication. We benchmark their performance against the off-the-shelf interior-point method Ipopt and against ADMM on infrastructure networks that contain up to 300,000 decision variables and constraints. Overall, we find that the proposed approaches solve problems as fast as Ipopt but with reduced communication. Moreover, we find that the proposed schemes achieve higher accuracy than ADMM approaches.
翻译:大规模基础设施网络的实时运作需要有可扩缩的优化能力。分解计划可以帮助实现可扩缩;传统的分解方法,如乘数和分布式牛顿计划交替方向法(ADMM)等典型分解方法,但是,通常要么是趋同缓慢,要么可能需要高度的沟通;在这项工作中,我们提出了解决大规模、强凝固的QP的新原始分解方法。这些方法具有全球趋同保证,需要有限的沟通。我们将其绩效与现成的内点方法Ipopt和含有多达30万个决定变量和制约因素的基础设施网络的ADMMM(ADM)对照。总体而言,我们发现,拟议的方法解决问题的速度与Ipott一样,但通信量减少。此外,我们发现,拟议的方案比ADMMM方针的精度更高。