Nat. Mach. Intell.:图神经网络(GNN),组合优化
<div style="color: black; text-align: left; margin-bottom: 10px;">
<div style="color: black; text-align: left; margin-bottom: 10px;">
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">组合优化Combinatorial optimization问题,<span style="color: black;">广泛</span>存在于科学和工业过程。现代深度学习deep learning工具,以前所未有的规模<span style="color: black;">处理</span>了这些问题,<span style="color: black;">不外</span>,结合统计<span style="color: black;">理学</span>学见解的统一框架,仍然很迫切。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">近期</span>,美国 亚马逊AWS 量子<span style="color: black;">处理</span><span style="color: black;">方法</span>实验室(Amazon Quantum So</span><span style="color: black;">lutions Lab) Martin J. A. Schuetz,J. Kyle Brubaker,Helmut G. Katzgraber,在Nature Machine Intelligence上<span style="color: black;">发帖</span>,<span style="color: black;">报告</span>演示了图神经网络graph neural networks<span style="color: black;">处理</span>组合优化问题。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">该<span style="color: black;">办法</span>广泛适用于二次无约束二元优化问题形式的典型NP-hard问题,如最大割maximum cut、最小顶点覆盖minimum vertex cover、最大独立集,以及多项式无约束二元优化问题形式的Ising Spin Glasses及其高阶推广<span style="color: black;">许多</span>方面。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">同期</span>,一种驰豫策略relaxation strategy应用于问题哈密顿量,以生成可微分的损失函数,该损失函数可用来训练图神经网络,并在无监督训练过程完成后,简单投影应用于整数变量。</span><span style="color: black;">标准最大割集和最大独立集问题的数值结果验证了这一<span style="color: black;">办法</span>。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">科研</span><span style="color: black;">发掘</span>,图神经网络优化器性能,相当或优于现有解算器,并且能够扩展到超过现有技术水平,以<span style="color: black;">处理</span><span style="color: black;">拥有</span>数百万变量的问题。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/T4YrUxuHnyXb22~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725642937&x-signature=IIqZlzXIgvVp%2FHI3K3fk04gsRiE%3D" style="width: 50%; margin-bottom: 20px;"></p>
<h1 style="color: black; text-align: left; margin-bottom: 10px;"><span style="color: black;">Combinatorial optimization with physics-inspired graph neural networks</span></h1><span style="color: black;">基于<span style="color: black;">理学</span>图神经网络的组合优化。</span>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/T4YrUyV7oYsmhE~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725642937&x-signature=I5WZTrHVPbR%2FfXDR5cwyURQ4Mhw%3D" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图1:组合优化图神经网络graph neural networks,GNN<span style="color: black;">办法</span>示意图。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/T4YrUzM1b8j4JS~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725642937&x-signature=t6XEijGM3E%2BU5Cqb0iD8A0abRQ8%3D" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图2:受<span style="color: black;">理学</span>学启发图神经网络GNN优化器的端到端工作流程图。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/T4YrV0Y2MT6DMT~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725642937&x-signature=kel23g8CzoMSLrffJmuqWQrc%2FAE%3D" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图3:n=100节点的随机3-正则图的最大割MaxCut示例解。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/T4YrV2N8JHrWvg~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725642937&x-signature=nNVngzwnv%2Fv4krQarkpcUqmjmxs%3D" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图4:最大割MaxCut数值结果。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-tjoges91tu/T4YrVYL6x5XSUa~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1725642937&x-signature=od2%2BLyxmaMlY37YG4QWIHEYQC5o%3D" style="width: 50%; margin-bottom: 20px;"></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">图5:最大独立集maximum independent set,MIS问题的数值结果。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">该项<span style="color: black;">科研</span>,提出并分析了一种多功能、可扩展的通用求解器,该求解器由图神经网络GNNS<span style="color: black;">供给</span>支持,并借鉴了统计<span style="color: black;">理学</span>学理念。该<span style="color: black;">办法</span>适用于任何k-局部伊辛k-local Ising模型,<span style="color: black;">包含</span>典型NP-hard组合优化问题,如最大割、最大团、最小顶点覆盖或最大独立集问题等66。从伊辛Ising公式出发,去掉决策变量上的完整性约束,并将驰豫策略relaxation strategy应用于问题哈密顿量,以生成可微损失函数,该损失函数对图神经网络GNN的节点<span style="color: black;">暗示</span>,进行无监督训练。<span style="color: black;">而后</span>,<span style="color: black;">针对</span>图中的<span style="color: black;">每一个</span>顶点,图神经网络GNN训练以生成软分配,预测属于两个类之一的可能性。为了找到与原始问题公式一致二元(双色)标记,采用简单的投影启发式projection heuristics 。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">总体而言,这种<span style="color: black;">办法</span>,<span style="color: black;">能够</span>与现有专用求解器竞争,例如设计用于<span style="color: black;">处理</span>最大切割问题的Goemans–Williamson算法,并有可能利用统计<span style="color: black;">理学</span>学的丰富工具箱,例如<span style="color: black;">包含</span>相变<span style="color: black;">科研</span>。在当前中等规模量子时代,该<span style="color: black;">办法</span><span style="color: black;">能够</span><span style="color: black;">做为</span>新兴量子技术(<span style="color: black;">包含</span>专用量子6和量子启发退火20)的广泛适用、可扩展基准,<span style="color: black;">同期</span>不受资源限制,<span style="color: black;">亦</span>不局限于QUB形式,相干伊辛机Ising machines<span style="color: black;">亦</span>是如此。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;">展望:</strong></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">1、为更好地理解图神经网络GNN,在组合优化环境中的局限性,进一步<span style="color: black;">科研</span>是有序的、系统的图神经网络GNN,与一大类优化问题的最先进解算器,进行基准测试,<span style="color: black;">同期</span>利用图神经网络GNN实现<span style="color: black;">全部</span>全局,<span style="color: black;">包含</span>例如Graphsage或Graph Attention Networks(GATS),以潜在地利用<span style="color: black;">重视</span>力机制,<span style="color: black;">提高</span>图神经网络GNN ansatz,该<span style="color: black;">重视</span>力机制,使得顶点能够在聚合<span style="color: black;">过程</span><span style="color: black;">时期</span>,对相邻<span style="color: black;">描述</span>neighbour representations进行加权。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">2、当在<span style="color: black;">设备</span>集群上,以小批量方式利用分布式训练时,所提出的图神经网络GNN<span style="color: black;">办法</span>,应该能够适应数亿个节点的问题规模,从而挑战几个现有解算器的能力。尽管<span style="color: black;">日前</span>实例,<span style="color: black;">重点</span><span style="color: black;">触及</span>随机初始化过程,初始节点嵌入问题,但在<span style="color: black;">将来</span>,<span style="color: black;">运用</span>预先训练的权重(迁移学习)<span style="color: black;">起步</span>训练过程,<span style="color: black;">能够</span><span style="color: black;">加强</span><span style="color: black;">处理</span>问题的时间。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">3、随机投影randomized projection<span style="color: black;">方法</span>,有望潜在地<span style="color: black;">加强</span>优化器性能,<span style="color: black;">或</span>贪心后处理greedy post-processing,<span style="color: black;">加强</span>这些策略,该后处理例程,<span style="color: black;">经过</span>一系列局部位翻转,<span style="color: black;">检测</span>局部最优性。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">4、该<span style="color: black;">办法</span><span style="color: black;">能够</span>推广到超图hypergraphsPUBO问题,其中超边hyper-edges <span style="color: black;">能够</span><span style="color: black;">包括</span>两个以上的节点,而不需要资源密集型度缩减<span style="color: black;">方法</span>,这与资源受限QUBO求解器相反。还有潜在的应用,涵盖了多体相互<span style="color: black;">功效</span>的现实世界优化问题,如在调度 scheduling 问题,或化学<span style="color: black;">发掘</span>问题。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">总之,<span style="color: black;">设备</span>学习、运筹学和<span style="color: black;">理学</span>学之间的交叉融合,开辟了许多新兴的<span style="color: black;">科研</span>方向,<span style="color: black;">最后</span><span style="color: black;">目的</span>是进一步<span style="color: black;">加强</span><span style="color: black;">处理</span>组合优化问题的能力。</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">文献链接:https://www.nature.com/articles/s42256-022-00468-6</span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">DOI</span><span style="color: black;">: https://doi.org/10.1038/s42256-022-00468-6</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">本文译自Nature。</p>
</div>
</div>
你的话深深触动了我,仿佛说出了我心里的声音。 你的见解独到,让我受益匪浅,非常感谢。 同意、说得对、没错、我也是这么想的等。 “沙发”(SF,第一个回帖的人) 楼主听话,多发外链好处多,快到碗里来!外链论坛 http://www.fok120.com/
页:
[1]