[论文解读] MemComputing Integer Linear Programming
本文提出了一种非算法性的内存计算方法,利用自组织代数电路求解整数线性规划(ILP)问题,通过标准硬件模拟内存计算机器。该方法在数分钟内求解了MIPLIB 2010库中长期未解的f2000问题——此前八年未知——相较于商用求解器展现出更优的可扩展性和解的质量。
Integer linear programming (ILP) encompasses a very important class of optimization problems that are of great interest to both academia and industry. Several algorithms are available that attempt to explore the solution space of this class efficiently, while requiring a reasonable compute time. However, although these algorithms have reached various degrees of success over the years, they still face considerable challenges when confronted with particularly hard problem instances, such as those of the MIPLIB 2010 library. In this work we propose a radically different non-algorithmic approach to ILP based on a novel physics-inspired computing paradigm: Memcomputing. This paradigm is based on digital (hence scalable) machines represented by appropriate electrical circuits with memory. These machines can be either built in hardware or, as we do here, their equations of motion can be efficiently simulated on our traditional computers. We first describe a new circuit architecture of memcomputing machines specifically designed to solve for the linear inequalities representing a general ILP problem. We call these self-organizing algebraic circuits, since they self-organize dynamically to satisfy the correct (algebraic) linear inequalities. We then show simulations of these machines using MATLAB running on a single core of a Xeon processor for several ILP benchmark problems taken from the MIPLIB 2010 library, and compare our results against a renowned commercial solver. We show that our approach is very efficient when dealing with these hard problems. In particular, we find within minutes feasible solutions for one of these hard problems (f2000 from MIPLIB 2010) whose feasibility, to the best of our knowledge, has remained unknown for the past eight years.
研究动机与目标
- 为解决整数线性规划(ILP)难题的长期挑战,特别是那些在现有算法下仍不可行或未解的MIPLIB 2010库中的问题。
- 开发一种非算法性、受物理启发的计算范式——内存计算——将问题嵌入动力学电路中,使其自然收敛至解。
- 设计并模拟自组织代数门(SOAGs),通过其内在动力学动态满足ILP问题的线性不等式约束。
- 在基准问题上展示该方法的效率与可扩展性,尤其在传统求解器失效或可扩展性差的情况下。
- 为MIPLIB 2010中的f2000 ILP实例提供一个可行解,该问题此前已持续八年未解。
提出的方法
- 该方法采用基于自组织代数门(SOAGs)的新型内存计算机器电路架构,这些门能动态自组织以满足给定ILP问题的线性不等式。
- 这些电路的动力学由一组常微分方程(ODEs)建模,其演化趋向于表示可行解的稳态。
- ODEs基于拉格朗日公式推导,通过罚函数强制约束,并确保收敛至可行解,且无混沌或周期轨道。
- 模拟使用MATLAB在单核Xeon处理器上进行,模拟数字内存计算机器的行为,无需量子硬件。
- 该方法与Gurobi 8.0通过式(2)定义的gap度量进行比较,衡量随时间推移的相对解质量。
- SOAGs的设计具有模块化和可扩展性,可使用标准电子元件实现硬件化。
实验结果
研究问题
- RQ1非算法性、受物理启发的计算范式(如内存计算)能否高效求解传统求解器难以处理的困难ILP问题?
- RQ2基于SOAG的电路的自组织行为是否能可靠收敛至复杂ILP实例的可行解?
- RQ3该方法在大规模问题(如MIPLIB 2010中的rmine系列)上能否保持高解质量和可扩展性?
- RQ4能否找到MIPLIB 2010中f2000问题的可行解,该问题此前已持续八年未解?
- RQ5与领先的商用求解器(Gurobi)相比,该内存计算方法在解质量与求解时间方面表现如何?
主要发现
- 该内存计算方法在数分钟内成功求解了MIPLIB 2010中f2000 ILP问题的可行解——据作者所知,该解此前已持续八年未被发现。
- 对于rmine基准系列,内存计算求解器(MemCPU)即使在大规模实例上也保持约1%的解差距,而Gurobi的性能显著下降,解差距达到数万个百分点。
- Gurobi的预处理时间随问题规模几乎呈四次方增长,尤其在较大的rmine实例上,导致其可扩展性差,且在时间限制内无法改善解差距。
- 内存计算方法展现出更优的可扩展性:尽管在小规模实例上较慢,但在大规模问题上保持了稳定性能,表明对问题规模增加具有鲁棒性。
- MemCPU的解质量在所有测试问题规模下均保持稳定且优异,对于大规模rmine实例,在3380秒计算后始终保持约1%的解差距。
- 所有解文件均以.sol格式公开发布于MemComputing, Inc.官网,可使用任何标准ILP求解器独立验证。
更好的研究,从现在开始
从阅读论文到最终审阅,大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成,并经人工编辑审核。