1: \begin{abstract}

2: Experimental design is a process of obtaining a product with target

3: property via experimentation. Bayesian optimization offers a sample-efficient

4: tool for experimental design when experiments are expensive. Often,

5: expert experimenters have 'hunches' about the behavior of the experimental

6: system, offering potentials to further improve the efficiency. In

7: this paper, we consider per-variable monotonic trend in the underlying

8: property that results in a unimodal trend in those variables for a

9: target value optimization. For example, sweetness of a candy is monotonic

10: to the sugar content. However, to obtain a target sweetness, the utility

11: of the sugar content becomes a unimodal function, which peaks at the

12: value giving the target sweetness and falls off both ways. In this

13: paper, we propose a novel method to solve such problems that achieves

14: two main objectives: a) the monotonicity information is used to the

15: fullest extent possible, whilst ensuring that b) the convergence guarantee

16: remains intact. This is achieved by a two-stage Gaussian process modeling,

17: where the first stage uses the monotonicity trend to model the underlying

18: property, and the second stage uses `virtual' samples, sampled from

19: the first, to model the target value optimization function. The process

20: is made theoretically consistent by adding appropriate adjustment

21: factor in the posterior computation, necessitated because of using

22: the `virtual' samples. The proposed method is evaluated through both

23: simulations and real world experimental design problems of a) new

24: short polymer fiber with the target length, and b) designing of a

25: new three dimensional porous scaffolding with a target porosity. In

26: all scenarios our method demonstrates faster convergence than the

27: basic Bayesian optimization approach not using such `hunches'.

28: \end{abstract}

29: