1: \begin{abstract}

2: In this paper, the problem of safe global maximization (it should

3: not be confused with robust optimization) of expensive noisy

4: black-box functions satisfying the Lipschitz condition is

5: considered. The notion ``safe" means that the objective function

6: $f(x)$ during optimization should not violate a ``safety" threshold,

7: for instance, a certain a priori given value $h$ in a maximization

8: problem. Thus, any new function evaluation (possibly corrupted by

9: noise) must be performed at ``safe points" only, namely, at points

10: $y$ for which it is known that the objective function $f(y) > h$.

11: The main difficulty here consists in the fact that the used

12: optimization algorithm should ensure that the safety constraint will

13: be satisfied at a point $y$ {\emph{before}} evaluation of $f(y)$

14: will be executed. Thus, it is required both to determine the safe

15: region $\Omega$ within the search domain~$D$ and to find the global

16: maximum within $\Omega$. An additional difficulty consists in the

17: fact that these problems should be solved in the presence of the

18: noise. This paper starts with a theoretical study of the problem and

19: it is shown that even though the objective function $f(x)$ satisfies

20: the Lipschitz condition, traditional Lipschitz minorants and

21: majorants cannot be used due to the presence of the noise. Then, a

22: $\delta$-Lipschitz framework and two algorithms using it are

23: proposed to solve the safe global maximization problem. The first

24: method determines the safe area within the search domain and the

25: second one executes the global maximization over the found safe

26: region. For both methods a number of theoretical results related to

27: their functioning and convergence is established. Finally, numerical

28: experiments confirming the reliability of the proposed procedures

29: are performed.

30: 

31: \vspace*{2mm} \textbf{Keywords:} {Safe global optimization;

32: expensive black-box functions; noise; Lipschitz condition; Machine

33: learning.}

34: 

35: \end{abstract}

36: