[{"data":1,"prerenderedAt":334},["ShallowReactive",2],{"content-query-T1y5b3KP4D":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":8,"date":9,"cover":10,"type":11,"body":12,"_type":328,"_id":329,"_source":330,"_file":331,"_stem":332,"_extension":333},"/technology-blogs/zh/555","zh",false,"","技术干货| 详解AI国际顶会NeurIPS 2020的黑盒优化竞赛冠军算法——HEBO算法","2021-05-18","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/18/e19d1e482a294a92815687f4bb19b2b2.png","technology-blogs",{"type":13,"children":14,"toc":325},"root",[15,23,32,62,69,74,81,86,93,100,105,110,117,135,140,147,159,166,183,188,195,200,207,219,226,238,245,250,259,264,273,278,287,292,303,308,313,318],{"type":16,"tag":17,"props":18,"children":20},"element","h1",{"id":19},"技术干货-详解ai国际顶会neurips-2020的黑盒优化竞赛冠军算法hebo算法",[21],{"type":22,"value":8},"text",{"type":16,"tag":24,"props":25,"children":26},"p",{},[27],{"type":16,"tag":28,"props":29,"children":31},"img",{"alt":7,"src":30},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/b0ff3501292148b79cd133d656f00d26.gif",[],{"type":16,"tag":24,"props":33,"children":34},{},[35,37,43,45,50,52,60],{"type":22,"value":36},"异方差演化贝叶斯优化（Heteroscedastic evolutionary bayesian optimisation, HEBO）算法是华为诺亚实验室提出的优化算法框架。该算法",{"type":16,"tag":38,"props":39,"children":40},"strong",{},[41],{"type":22,"value":42},"击败NVIDIA",{"type":22,"value":44},"，",{"type":16,"tag":38,"props":46,"children":47},{},[48],{"type":22,"value":49},"IBM",{"type":22,"value":51},"等赢得了AI国际顶会NeurIPS 2020的黑盒优化竞赛冠军（",{"type":16,"tag":53,"props":54,"children":58},"a",{"href":55,"rel":56},"https://bbochallenge.com/leaderboard",[57],"nofollow",[59],{"type":22,"value":55},{"type":22,"value":61},"）。该算法提出了对代理模型和最大化获取函数的非常规修改，并显示了这种组合优于Bayesmark包提供的所有基线。",{"type":16,"tag":24,"props":63,"children":64},{},[65],{"type":16,"tag":28,"props":66,"children":68},{"alt":7,"src":67},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/74ab2ef974124ee79cf6af02a32baebc.png",[],{"type":16,"tag":24,"props":70,"children":71},{},[72],{"type":22,"value":73},"黑盒优化问题中的目标函数的表达式未知，只能根据离散的自变量取值得到对应的目标函数值来更新取值。除了AutoML和超参数优化以外，黑盒优化在分子设计、材料科学、电厂调试、数据中心冷却、芯片设计等问题上也有着大量应用。",{"type":16,"tag":24,"props":75,"children":76},{},[77],{"type":16,"tag":28,"props":78,"children":80},{"alt":7,"src":79},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/e7f9a5b947854c0ebda1d5afaa64016a.jpg",[],{"type":16,"tag":24,"props":82,"children":83},{},[84],{"type":22,"value":85},"一般黑盒优化算法需要经过特定步骤实现优化迭代过程，如下图所示：",{"type":16,"tag":24,"props":87,"children":88},{},[89],{"type":16,"tag":28,"props":90,"children":92},{"alt":7,"src":91},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/54f9e058e4384e3c95b13ed1177f60f9.jpg",[],{"type":16,"tag":24,"props":94,"children":95},{},[96],{"type":16,"tag":28,"props":97,"children":99},{"alt":7,"src":98},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/ef1209a22e9647c69330cfa7427147c2.png",[],{"type":16,"tag":24,"props":101,"children":102},{},[103],{"type":22,"value":104},"在研究过程中，研究者发现黑盒优化中最关键的两个问题是：",{"type":16,"tag":24,"props":106,"children":107},{},[108],{"type":22,"value":109},"在步骤1拟合代理模型阶段，通常会用高斯模型来作为代理模型，首先获得输入数据集，假设一个高斯模型为代理模型，用数据去拟合这个代理模型，如下图所示：",{"type":16,"tag":24,"props":111,"children":112},{},[113],{"type":16,"tag":28,"props":114,"children":116},{"alt":7,"src":115},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/9598a638cc76432482b368f64080499c.jpg",[],{"type":16,"tag":24,"props":118,"children":119},{},[120,122,127,128,133],{"type":22,"value":121},"但是",{"type":16,"tag":38,"props":123,"children":124},{},[125],{"type":22,"value":126},"输入的数据通常是具有噪声的",{"type":22,"value":44},{"type":16,"tag":38,"props":129,"children":130},{},[131],{"type":22,"value":132},"不符合高斯分布",{"type":22,"value":134},"。",{"type":16,"tag":24,"props":136,"children":137},{},[138],{"type":22,"value":139},"在步骤2最大化采集函数阶段，贝叶斯优化中通用的采集函数有EI（Expected Improvement）、PI（Probability of improvement）、UCB（Upper Confidence Bound），采集函数如下图所示：",{"type":16,"tag":24,"props":141,"children":142},{},[143],{"type":16,"tag":28,"props":144,"children":146},{"alt":7,"src":145},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/3eb16562958f4e1b8691193a25e9fa66.jpg",[],{"type":16,"tag":24,"props":148,"children":149},{},[150,152,157],{"type":22,"value":151},"他们都能帮助下一步在黑盒函数中评估哪些点，但是目前还",{"type":16,"tag":38,"props":153,"children":154},{},[155],{"type":22,"value":156},"不清楚在不同任务时哪个最有效",{"type":22,"value":158},"，那如何根据不同的任务选择采集函数。",{"type":16,"tag":24,"props":160,"children":161},{},[162],{"type":16,"tag":28,"props":163,"children":165},{"alt":7,"src":164},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/0aba63af1d3a4bfe8a8b6f4f662e0c05.png",[],{"type":16,"tag":24,"props":167,"children":168},{},[169,171,176,177,182],{"type":22,"value":170},"为了解决问题一，HEBO在采用贝叶斯优化中常用的高斯过程模型（Gaussian process, GP）作为代理模型的同时，还",{"type":16,"tag":38,"props":172,"children":173},{},[174],{"type":22,"value":175},"解决非平稳目标问题",{"type":22,"value":44},{"type":16,"tag":38,"props":178,"children":179},{},[180],{"type":22,"value":181},"引入输入扭曲以及power变换",{"type":22,"value":134},{"type":16,"tag":24,"props":184,"children":185},{},[186],{"type":22,"value":187},"对每个输入使用Kumaraswamy分布的累积分布函数，公式和效果图如下所示下图所示：",{"type":16,"tag":24,"props":189,"children":190},{},[191],{"type":16,"tag":28,"props":192,"children":194},{"alt":7,"src":193},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/f369dde8ce2440119db8dc191c4821fa.jpg",[],{"type":16,"tag":24,"props":196,"children":197},{},[198],{"type":22,"value":199},"对输出的变量使用power变换，使得输出的数据更好地近似于高斯分布，公式和效果图如下图所示：",{"type":16,"tag":24,"props":201,"children":202},{},[203],{"type":16,"tag":28,"props":204,"children":206},{"alt":7,"src":205},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/1351888ff335472eac8c81e9faf881c4.jpg",[],{"type":16,"tag":24,"props":208,"children":209},{},[210,212,217],{"type":22,"value":211},"为了解决问题二，研究者提出了MACE（Multi-objective ACquisition function Ensemble），将贝叶斯优化中通用的EI、PI、UCB采集函数",{"type":16,"tag":38,"props":213,"children":214},{},[215],{"type":22,"value":216},"进行聚合",{"type":22,"value":218},"，在这三种采集函数的帕累托前沿上进行采样，一方面避免单一采集函数的通用性问题，另一方面支持并行探索。如下图所示：",{"type":16,"tag":24,"props":220,"children":221},{},[222],{"type":16,"tag":28,"props":223,"children":225},{"alt":7,"src":224},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/7adf57dca48446dc937da121f057c30e.jpg",[],{"type":16,"tag":24,"props":227,"children":228},{},[229,231,236],{"type":22,"value":230},"HEBO相对于以前最先进的TuRBO（发表于NeurIPS 2019）算法分数",{"type":16,"tag":38,"props":232,"children":233},{},[234],{"type":22,"value":235},"提高了5.17%",{"type":22,"value":237},"，是与前5名竞争对手之间差异最大的算法，所以它以非常大的优势获胜。下面是比赛结果的截图：",{"type":16,"tag":24,"props":239,"children":240},{},[241],{"type":16,"tag":28,"props":242,"children":244},{"alt":7,"src":243},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/cdb5ba1712124861bbd4eed67e5c9585.jpg",[],{"type":16,"tag":24,"props":246,"children":247},{},[248],{"type":22,"value":249},"目前HEBO-MindSpore已经在github上开源：",{"type":16,"tag":24,"props":251,"children":252},{},[253],{"type":16,"tag":53,"props":254,"children":257},{"href":255,"rel":256},"https://github.com/huawei-noah/noah-research/tree/mindspore/HEBO_mindspore",[57],[258],{"type":22,"value":255},{"type":16,"tag":24,"props":260,"children":261},{},[262],{"type":22,"value":263},"也可以直接在pypi网站下载安装：",{"type":16,"tag":24,"props":265,"children":266},{},[267],{"type":16,"tag":53,"props":268,"children":271},{"href":269,"rel":270},"https://pypi.org/project/HEBO-mindspore/",[57],[272],{"type":22,"value":269},{"type":16,"tag":24,"props":274,"children":275},{},[276],{"type":22,"value":277},"你可以在这里下载文章链接：",{"type":16,"tag":24,"props":279,"children":280},{},[281],{"type":16,"tag":53,"props":282,"children":285},{"href":283,"rel":284},"https://arxiv.org/abs/2012.03826",[57],[286],{"type":22,"value":283},{"type":16,"tag":24,"props":288,"children":289},{},[290],{"type":22,"value":291},"MindSpore官方资料",{"type":16,"tag":24,"props":293,"children":294},{},[295,297],{"type":22,"value":296},"GitHub : ",{"type":16,"tag":53,"props":298,"children":301},{"href":299,"rel":300},"https://github.com/mindspore-ai/mindspore",[57],[302],{"type":22,"value":299},{"type":16,"tag":24,"props":304,"children":305},{},[306],{"type":22,"value":307},"Gitee:[https : //gitee.com/mindspore/mindspore](https : //gitee.com/mindspore/mindspore)",{"type":16,"tag":24,"props":309,"children":310},{},[311],{"type":22,"value":312},"官方QQ群 : 871543426",{"type":16,"tag":24,"props":314,"children":315},{},[316],{"type":22,"value":317},"长按下方二维码加入MindSpore项目↓",{"type":16,"tag":24,"props":319,"children":320},{},[321],{"type":16,"tag":28,"props":322,"children":324},{"alt":7,"src":323},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2021/05/17/a4632ed920de4395b2f38692c5dca692.jpg",[],{"title":7,"searchDepth":326,"depth":326,"links":327},4,[],"markdown","content:technology-blogs:zh:555.md","content","technology-blogs/zh/555.md","technology-blogs/zh/555","md",1776506137754]