1"use strict";(self.webpackChunk=self.webpackChunk||[]).push([[48],{16208(e,t,n){n.r(t),n.d(t,{assets:()=>d,contentTitle:()=>l,default:()=>h,frontMatter:()=>r,metadata:()=>i,toc:()=>c});const i=JSON.parse('{"id":"intro-to-ae","title":"Introduction to Adaptive Experimentation","description":"In engineering tasks we often encounter so-called \\"black box\\" optimization","source":"@site/versioned_docs/version-1.1.2/intro-to-ae.mdx","sourceDirName":".","slug":"/intro-to-ae","permalink":"/docs/1.1.2/intro-to-ae","draft":false,"unlisted":false,"tags":[],"version":"1.1.2","lastUpdatedBy":"github-actions[bot]","lastUpdatedAt":1757446870000,"frontMatter":{"id":"intro-to-ae","title":"Introduction to Adaptive Experimentation"},"sidebar":"docs","previous":{"title":"Installation","permalink":"/docs/1.1.2/installation"},"next":{"title":"Introduction to Bayesian Optimization","permalink":"/docs/1.1.2/intro-to-bo"}}');var o=n(74848),a=n(28453);const s=n.p+"assets/images/traditional_vs_adaptive-435c18766f0156ac7ec86dbda07f8f83.png",r={id:"intro-to-ae",title:"Introduction to Adaptive Experimentation"},l="Introduction to Adaptive Experimentation",d={},c=[];function p(e){const t={em:"em",h1:"h1",header:"header",li:"li",ol:"ol",p:"p",strong:"strong",ul:"ul",...(0,a.R)(),...e.components};return(0,o.jsxs)(o.Fragment,{children:[(0,o.jsx)(t.header,{children:(0,o.jsx)(t.h1,{id:"introduction-to-adaptive-experimentation",children:"Introduction to Adaptive Experimentation"})}),"\n",(0,o.jsx)(t.p,{children:'In engineering tasks we often encounter so-called "black box" optimization\nproblems, situations where the relationship between inputs and outputs of a\nsystem is not known in advance. In these scenarios practitioners must tune\nparameters using many time- and/or resource-consuming trials. For example:'}),"\n",(0,o.jsxs)(t.ul,{children:["\n",(0,o.jsx)(t.li,{children:"Machine learning engineers and researchers may have neural network\narchitectures and training procedures that may depend on numerical\nhyperparameters, such as learning rate, number of embedding layers or widths,\ndata weights, data augmentation choices, etc. One often seeks to understand\nand/or optimize with respect to these tradeoffs ."}),"\n",(0,o.jsx)(t.li,{children:"Materials scientists may seek to find the composition and heat treatment\nparameters that maximize strength for an alloy."}),"\n",(0,o.jsx)(t.li,{children:"Chemists may seek to find the synthesis path for a molecule that is likely to\nbe a good drug candidate for a disease."}),"\n"]}),"\n",(0,o.jsx)(t.p,{children:"Adaptive experimentation is an approach to solving these problems efficiently by\nactively proposing new trials to run as additional data is received. Adaptive\nexperimentation is able to explore large configuration spaces with limited\nresources through the use of specialized models and optimization algorithms."}),"\n",(0,o.jsx)(t.p,{children:"The basic adaptive experimentation flow works as follows:"}),"\n",(0,o.jsxs)(t.ol,{children:["\n",(0,o.jsxs)(t.li,{children:[(0,o.jsx)(t.strong,{children:"Configure"})," your optimization experiment, defining the space of values to\nsearch over, objective(s), constraints, etc."]}),"\n",(0,o.jsxs)(t.li,{children:[(0,o.jsx)(t.strong,{children:"Suggest"})," new trials, to be evaluated one at a time or in a parallel (a\n\u201cbatch\u201d)"]}),"\n",(0,o.jsxs)(t.li,{children:[(0,o.jsx)(t.strong,{children:"Evaluate"})," the suggested trials by executing the black box function and\nreporting the results back to the optimization algorithm"]}),"\n",(0,o.jsxs)(t.li,{children:[(0,o.jsx)(t.strong,{children:"Repeat"})," steps 2 and 3 until a stopping condition is met or the evaluation\nbudget is exhausted"]}),"\n"]}),"\n",(0,o.jsx)("center",{children:(0,o.jsx)("img",{src:s,alt:"Traditional vs. Adaptive design",width:"50%"})}),"\n",(0,o.jsxs)(t.p,{children:["Bayesian optimization, one of the most effective forms of adaptive\nexperimentation, intelligently balances tradeoffs between exploration (learning\nhow new parameterizations perform) and exploitation (refining parameterizations\npreviously observed to be good). To achieve this, Bayesian optimization\nutilizes a ",(0,o.jsx)(t.em,{children:"surrogate model"})," (most commonly, a Gaussian process) to predict the\nbehavior of the \u201cblack box\u201d at any given input configuration (parameterization,\nwh
1atever we want to call it). An ",(0,o.jsx)(t.em,{children:"acquisition function"})," utilizes predictions from\nthis model to identify new parameterizations to identify promising candidates to\nevaluate."]})]})}function h(e={}){const{wrapper:t}={...(0,a.R)(),...e.components};return t?(0,o.jsx)(t,{...e,children:(0,o.jsx)(p,{...e})}):p(e)}},28453(e,t,n){n.d(t,{R:()=>s,x:()=>r});var i=n(96540);const o={},a=i.createContext(o);function s(e){const t=i.useContext(a);return i.useMemo(function(){return"function"==typeof e?e(t):{...t,...e}},[t,e])}function r(e){let t;return t=e.disableParentContext?"function"==typeof e.components?e.components(o):e.components||o:s(e.components),i.createElement(a.Provider,{value:t},e.children)}}}]);
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.