贝叶斯优化

BAYESIAN OPTIMIZATION 贝叶斯优化

Consider the following problem of finding a global minimizer (or maximizer) of an unknown objective function：
$${\mathbf{x}}^{\ast }=\arg \underset{\mathbf{x}\in \mathcal{X}}{\min }f(\mathbf{x})$$

注意，$f$ 可能是非凸的，梯度信息是不可用的。
超参数搜索空间 $\mathcal{X}$。

由于目标函数是未知的，贝叶斯策略是把它当作一个随机函数，并放置一个先验（prior）。

• 先验 (prior) 是关于函数行为的置信度（belief）。
• 在收集到可以作为数据的函数评估后，对先验信息进行更新，形成目标函数的后验分布。
• 再用后验分布构造一个采集（acquisition）函数，确定下一个查询点。

---

widely used Gaussian prior

$$f({\mathbf{x}}_{1:t})=[f({\mathbf{x}}_1),\cdots ,f({\mathbf{x}}_t){]}^T\sim \mathcal{N}(0,{\mathbf{K}}_t)$$

• kernel matrix $t\times t$ and ${\mathbf{K}}_t(i,j)=k(\Vert {\mathbf{x}}_i-{\mathbf{x}}_j\Vert )$

Two popular kernels

• squared exponential (SE) kernel （径向基函数核，也被称为高斯核或平方指数核）
• the Matérn kernel（马特恩协方差函数）

• $\Gamma (\ast )$ Gamma function
• $B_{\nu }(\ast )$ $\nu$-th order Bessel function
• $h$ hyper-parameter

In Bayesian optimization, at the $t$-th iteration

• samples ${\mathcal{D}}_{1:t}={\left\{({\mathbf{x}}_i,f({\mathbf{x}}_i))\right\}}_{i=1}^t$
• at the next query point, inferring the value of $f({\mathbf{x}}_{i+1})$

根据高斯先验假设，

• ${\mathbf{k}}_{t+1}={\left[k(\Vert {\mathbf{x}}_{t+1}-{\mathbf{x}}_1\Vert ),\cdots ,k(\Vert {\mathbf{x}}_{t+1}-{\mathbf{x}}_t\Vert )\right]}^T$

• 由于 $\left[\begin{array}{c}f({\mathbf{x}}_{1:t})\\ f({\mathbf{x}}_{t+1})\end{array}\right]$ 是联合正态的，条件分布 $f({\mathbf{x}}_{t+1})|f({\mathbf{x}}_{1:t})$ 也一定是正态的，我们可以使用该条件分布的均值和方差的标准公式

---

Reduce the complexity

注意，(6) 和 (7) 的计算复杂度很高，因为矩阵和向量的维数随 t 增长。

acquisition function

Now that we have a model of the function and its uncertainty, we will use this to choose which point to sample next.

• The acquisition function takes the mean and variance at each point on the function and computes a value that indicates how desirable it is to sample next at this position.

• A good acquisition function should trade off exploration and exploitation.

four popular acquisition functions:

• the upper confidence bound
• expected improvement
• probability of improvement
• Thompson sampling

Upper confidence bound

Direct balance between exploration and exploitation:
This acquisition function is defined as:

Expected Improvement

• Perhaps the most used acquisition.
• It is too greedy in some problems. It is possible to make more explorative adding a ’explorative’ parameter
  

https://www.cnblogs.com/marsggbo/p/9866764.html

https://www.cnblogs.com/marsggbo/p/9866764.html