1: \begin{abstract}

2: We present an improved stochastic trust-region algorithm with adaptive sampling (ASTRO-DF) with two noticeable features. First, each iteration of the algorithm constructs a local model with a diagonal Hessian using coordinate basis points to find the next incumbent. This provides sufficiently accurate gradient and Hessian with $\mcO(d)$ design points and linear algebra operations as opposed to $\mcO(d^2)$ in the original algorithm. Second, a direct search feature increases the probability of success and improvement in the optimization, keeping the trust-region size to become too small and hence computation budget exhaustion too early. These refinements have a significant effect in finite-time performance, notwithstanding almost sure convergence and canonical iteration complexity of $\mcO(\epsilon^{-2})$. Numerical results on SimOpt library problems and solvers supports the gained advantages of ASTRO-DF refined with direct coordinate search.\\

3: 

4: %\textcolor[rgb]{0.00,0.07,1.00}{We strongly encourage authors to address the following three questions in their \textbf{abstract}, preferably following the order shown: 1 Research problem statement: what is the research problem to be addressed? (2) Methods and results: how do the authors address the research problem and what are the main results? (3) Insights and implications:  What have the authors learned (as opposed to what they did, which is covered in point (2)) from conducting this research? What is the knowledge gained and why does it matter? The abstract should be written in \textbf{a single paragraph}.}.\\

5: 

6: %We thank you for your attention to these details.

7: 	\end{abstract}

8: