SOP: parallel surrogate global optimization with Pareto center selection for computationally expensive single objective problems
 1.5k Downloads
 2 Citations
Abstract
This paper presents a parallel surrogatebased global optimization method for computationally expensive objective functions that is more effective for larger numbers of processors. To reach this goal, we integrated concepts from multiobjective optimization and tabu search into, single objective, surrogate optimization. Our proposed derivativefree algorithm, called SOP, uses nondominated sorting of points for which the expensive function has been previously evaluated. The two objectives are the expensive function value of the point and the minimum distance of the point to previously evaluated points. Based on the results of nondominated sorting, P points from the sorted fronts are selected as centers from which many candidate points are generated by random perturbations. Based on surrogate approximation, the best candidate point is subsequently selected for expensive evaluation for each of the P centers, with simultaneous computation on P processors. Centers that previously did not generate good solutions are tabu with a given tenure. We show almost sure convergence of this algorithm under some conditions. The performance of SOP is compared with two RBF based methods. The test results show that SOP is an efficient method that can reduce time required to find a good near optimal solution. In a number of cases the efficiency of SOP is so good that SOP with 8 processors found an accurate answer in less wallclock time than the other algorithms did with 32 processors.
Keywords
Radial basis functions Tabu Computationally expensive Blackbox Simulation optimization Response surface Metamodel1 Introduction
Realworld applications in various fields, such as physics, engineering, or economics, often have a simulation model which is multimodal, computationally expensive, and blackbox. “Blackbox” implies that many mathematical characteristics are not known, including derivatives or number of local minima. Many existing optimization methods for blackbox functions such as genetic algorithm, simulated annealing, or particle swarm are not suitable for this type of problem due to the large number of objective function evaluations that these methods generally require.
One approach for dealing with this type of problem is to use a surrogate model (alternatively called metamodel or response surface) to approximate the objective function. Response surface based optimization methods start by building a (computationally inexpensive) surrogate surface, which is then used to iteratively select new points for the expensive function evaluation. The surrogate surface is updated in each iteration.
The purpose of this research is to develop a new way to solve surrogate optimization problems for computationally expensive functions in parallel. Previous efforts (mostly serial) have involved generating candidate points by normal perturbations around one center point (usually the best solution found so far), by uniform sampling in the domain, or by using an optimization method on the surrogate to find the point that satisfies some criterion. In this work, we use a different approach involving nondominated sorting on previously evaluated points to select multiple centers, which are points on the sorted fronts. Hence, we are using concepts from multiobjective optimization for single objective optimization of computationally expensive functions.
1.1 Literature review
Many authors have demonstrated the effectiveness of using response surfaces for optimization of computationally expensive problems within a limited number of function evaluations. Jones et al. [10] used kriging as a basis function to develop a global optimization method, Efficient Global Optimization (EGO), where the next point to evaluate is obtained by maximizing an expected improvement, balancing the response surface value with the uncertainty of the surface. Huang et al. [9] extended Jones’ EGO algorithm and developed a global optimization method for noisy cost functions. Booker et al. [3] also used kriging surface to speed up the pattern search algorithms. Gutmann [6] built the response surface with radial basis functions where the next point to evaluate is obtained by minimizing the bumpiness of the interpolation surface. Sóbester et al. [23] used Gaussian radial basis functions and proposed weighted expected improvement to control the balance of exploitation (of the response surface) and exploration (of the decision variable space) to select the next evaluation point. Regis and Shoemaker [18] also used radial basis functions in the Metric Stochastic Response Surface algorithm (also known as the Stochastic RBF algorithm), which is a global optimization algorithm wherein the next point to evaluate is chosen by randomly generating a large number of candidate points and selecting the point that minimizes a weighted score of response surface predictions and a distance metric. Optimization of computationally expensive problems is still an active field of research as can be seen by several workshops devoted to this subject. More recent papers on this subject include e.g. [11, 12, 15].
Due to the pervasiveness of parallel computing, there is a need to develop surrogate algorithms that select and evaluate multiple points in each iteration to reduce the wallclock time (which is proportional to the number of optimization iterations). For example, Sóbester et al. [22] developed a parallel version of EGO. Several local maximum points of the expected improvement function are selected for the expensive evaluations in parallel. In 2009, Regis and Shoemaker proposed a parallel version of the Stochastic RBF algorithm [19]. In each iteration, a fixed number of points are selected for doing the expensive function evaluations. The selection is done sequentially and based on the weighted score of (1) the surrogate value, and (2) the minimum distance from previously evaluated points and previously selected points within that parallel iteration, until the desired number of points are selected. The expensive function evaluations at the selected points are done simultaneously. The experimental results showed that the algorithm is very efficient compared to their previous methods. As a counterpart of EGO, Viana et al. [25] proposed MSEGO that is based on the idea of maximizing the expected improvement (EI) functions, but instead of using just one surrogate as in EGO, multiple surrogates are used. Since different EI functions are obtained for different surrogates, multiple points can be selected per iteration. Although MSEGO was shown to reduce the number of iterations, they found that the numerical convergence rate did not scale up with the number of points selected in each iteration for evaluation.
There is an inherent tradeoff between exploration and exploitation in surrogatebased optimization methods. Recently, Bischl et al. [2] attempted at analyzing this tradeoff and proposed MOIMBO which is a parallel kriging based optimization algorithm that uses a multiobjective infill criterion that rewards the diversity and the expected improvement for selecting the next expensive evaluation points. Many versions of MOIMBO were proposed based on various multiobjective infill criteria: mean of the surrogate model, model uncertainty, expected improvement, distance to the nearest neighbor, and the distance to the nearest better neighbor. Evolutionary optimization was used to handle the embedded multiobjective optimization problem. The overall test results suggested that the version that used a combination of mean, model uncertainty, and nearest neighbor worked best.
1.2 Differences between SOP and previous algorithms
The major difference between our algorithm and other existing parallel optimization algorithms is the use of a Pareto nondominated sorting technique to select P distinct evaluated points whose objective function values are small and that are far away from other evaluated points, which will then serve as centers. The selected centers are subsequently used with the addition of random perturbations for generating a set of candidate points from which the next function evaluation points are chosen and evaluated. In addition the selection of the P points is subject to tabu constraints. This contrasts with the approach Regis and Shoemaker [19] used to generate the P points for parallel expensive function evaluations. In [19], all the P points are obtained from the same center that is the best point found so far. Selecting the P points from different centers (as is done in this work) allows the algorithm to search more globally simultaneously in each iteration, which is especially helpful as the number of processors increases. The concept of nondominated sorting has been widely used in multiobjective optimization [1, 4, 26]. SOP considers the tradeoff between exploration and exploitation in single objective optimization as a biobjective optimization problem and incorporates nondominated sorting into the algorithm framework. The method proposed in [2] is also based on multiobjective optimization. However, their embedded multiobjective optimization problem was solved on the surrogate while in our approach, a biobjective optimization problem is solved on previously evaluated expensive function evaluation points.
We found no journal papers on surrogate global optimization that select a large number of evaluation points in each iteration in order to efficiently use a large number of processors. For example, the maximum number of points used in [2, 19], and [25] were 5, 8, and 10 points, respectively. On the other hand, SOP can do many expensive objective function evaluations simultaneously. (SOP was tested on as many as 64 points per parallel iteration.) SOP thus has the potential to greatly reduce wallclock time.
The structure of this paper is as follows: The new algorithm, SOP, is described in Sect. 2, and its theoretical properties are described in Theorem 1. In Sect. 3, we illustrate the performance of SOP and compare algorithms on a number of test functions as well as a groundwater bioremediation problem. Lastly, we give our concluding remarks in Sect. 4.
2 Algorithm description
Surrogate Optimization with Pareto Selection (SOP) is a parallel surrogate based algorithm where simultaneous surrogateassisted candidate searches are performed around selected evaluated points (referred to as centers in subsequent discussions). The search mechanism of the algorithm incorporates biobjective optimization, where the conflicting objectives focus the search to achieve a balance between exploration and exploitation, tabu search, and surrogate assisted candidate search. The motivation for using this combination of methods is so we can generate P points (where P can be large) for evaluation in parallel that will provide useful information for the search. Hence, these points need to be efficiently distributed and among other factors should not be too close to each other.
A synchronous Masterslave parallel framework is employed for simultaneous surrogateassisted candidate search. The algorithm is adaptive and learns from the search results of prior iterations.
2.1 General algorithm framework
The algorithm initialization phase (Step 1) starts by evaluating the expensive f(x) on \(n_0\) initial points from the decision space. The initial point may be selected via any experimental design method, e.g. Latin hypercube sampling. The expensive points are subsequently evaluated before initiation of the iterative loop.
Step 2 of the algorithm corresponds to the iterative loop, which has four subcomponents, namely (2.1) Fit/refit surrogate model, (2.2) Select the P centers, (2.3) Parallel candidate search, and (2.4) Adaptive learning.
In Step 2.1, all previously evaluated points and their corresponding function values are used to fit a surrogate model. Steps 2.2, 2.3 and 2.4 constitute the core components that define the search mechanics of SOP by maintaining a balance between exploration and exploitation, and incorporating the power of parallelism for improving efficiency of the surrogateassisted candidate search. Each of these steps will be discussed in more detail later.
At the end of the iterative loop (Step 3), the algorithm terminates after a prespecified number of iterations (or number of function evaluations) and returns the best solution found so far.
2.2 Step 2.2 nondominated sorting and P center selection
Given that P processors are available, Step 2.2 of SOP selects P points from the set of already evaluated points as center points for parallel surrogateassisted candidate search. The centers are selected by using nondominated sorting (an idea from multiobjective optimization) on points where the expensive f(x) has been evaluated. This approach uses two objective functions to find a diverse set of points that are likely to provide good function values and be diverse enough to improve the surrogate surface when many points are being evaluated simultaneously by parallel processors. While previous methods, such as Parallel Stochastic RBF, define the best solution found in all previous iterations as a center for perturbations, we instead create as many centers for perturbations as there are processors to facilitate improved parallel performance.
Let \(S^{(n)}\) be the set of already evaluated points after n algorithm iterations. SOP considers the challenge of balancing the tradeoff between exploration and exploitation as a biobjective optimization problem.
The objective function corresponding to exploitation is simply the objective function of the expensive optimization problem and is referred to as \(F_{1}(x)\). The second objective \(F_{2}(x)\) focusing on exploration is the minimum distance of an evaluated point, \(x \in S^{(n)}\), from all other evaluated points, \(S^{(n)}\setminus \{x\}\). A large value of minimum distance intuitively indicates that a potential center point is in a sparsely explored region of the decision space where the accuracy of the current solution can possibly be improved with candidate search in that region. Thus, to explore previously unexplored regions, a large value of \(F_{2}(x)\) is desirable. Maximizing \(F_{2}(x)\) is the same as minimizing its negative, so the second objective \(F_{2}(x)\) is defined as the negative of the minimum distance to the set of already evaluated points.
Given that SOP considers the process of selection of centers as the biobjective problem defined above, P points are selected as centers from the set of evaluated points. The process for selection of P centers is depicted in Fig. 2.
The selection process initiates by ranking all evaluated points \(\left( x_i, f\left( x_i\right) \right) \) according to the twotier ranking system.^{2} This corresponds to the upper left box in Fig. 2. The first tier of ranking is based on the concept of nondominated sorting method [5] on the biobjective problem defined in Eq. 2, which divides the evaluated set into mutually exclusive subsets which are referred to as fronts, where each front has a unique rank. The first rank (Pareto front) is given to the nondominated solutions in the evaluated set. These solutions are then removed, and the nondominated solutions identified in the remaining set are given the second rank. This process continues until all solutions are sorted into fronts (see Fig. 3b).
Since multiple solutions may be on the same front, we introduce the second tier of ranking to differentiate between these solutions. The solutions on the same front are ranked according to their objective function values f(x) (best to worst). Hence, the best evaluated point found so far has the best rank (since it lies on the first front and has the best objective function value). The secondary ranking tends to focus more on exploitation than exploration. We will denote by Ranked List the set (list) of evaluated points sorted according to this twotier ranking system.
The first selection criteria is that a point in the Ranked List is only selected as a center if it is not in the Tabu List. The Tabu List is a subset of evaluated points and contains points which, after being chosen as center points in \(N_{{\text {fail}}}\) algorithm iterations, did not induce an improvement in the explorationexploitation tradeoff. Details of the process of how points are added and removed from the Tabu List are later discussed in Step 2.4 of SOP. The premise behind maintaining a Tabu List is to induce adaptive learning within the SOP search mechanism that could ensure that points which do not induce improvement in the explorationexploitation tradeoff after repeated candidate searches are identified and subsequently removed from consideration as search centers in subsequent algorithm iterations. This selection condition is referred to as the Tabu Rule.
The second selection criteria is that a point is only selected as a center if its distance from every point that is already selected as a center is greater than the candidate search neighborhood sampling radius of the selected center. The candidate search neighborhood sampling radius for all evaluated points is adaptive, with an initial value equal to \(r_{{\text {int}}}\). The effect of changing this parameter was examined in [20]. Their results suggest that on more complicated surfaces a relatively large initial search radius is more effective since it allows the algorithm to explore a large region initially before focusing the search on a smaller region. Step 2.4 of SOP explains how the value of the candidate search neighborhood of a center point adapts. This selection criteria maintains the notion of exploration within the center selection framework by not selecting a point as a center if it lies within the candidate search sampling radius of an already selected center of the current algorithm iteration. This selection condition is referred to as the Radius Rule.
At the end of this process, we will have obtained \(C=\{c_1, \dots ,c_P\}\), the set of P selected centers.
Figure 3 illustrates an example of the center selection process. First the points in \(S^{(n)}\) are sorted according to the twotier ranking system (Fig. 3b). In this example, four center points (\(P=4\)) are selected sequentially from the Ranked List. First, the point A1 on the first front is selected because A1 has the minimum value of \(F_1(x)\). Since A2 is within the candidate search neighborhood radius of A1, and A3 is in the Tabu List, neither of these two points are selected as centers. We move on to the second front and the next point that will be selected is therefore B1 (since it has the lowest objective function \(F_1(x)\) value on this front). Continuing with the selection process, the four points that satisfy both the Radius Rule and the Tabu Rule are A1, B1, B3, and C2, and are thus selected as centers, and the process stops.
In the special case where all points in the Ranked List have already been examined but the number of selected centers is less than P (the number of centers needed), we reexamine points in the Ranked List with only the Radius Rule imposed. If the number of centers selected is still less than P, the next \(Pi\) centers, \(c_{i+h},\, h=1,..,\, Pi\), will be selected iteratively by cycling through the set of already selected centers \(\left\{ c_{1},...,\, c_{i}\right\} \), i.e. \(c_{qi+l}=c_{l}\) for \(q\ge 1\) and \(1\le l\le i\), until a total of P centers are selected. This is because good solutions often lie in a very small neighborhood. Therefore, it may be beneficial to focus the local candidate search around those promising centers.
2.3 Step 2.3 candidate search
Once P center points have been selected, in Step 2.3 of SOP, simultaneous surrogateassisted candidate searches are performed around each center for selection of new points for expensive evaluation. For every center point, one new point is selected for expensive evaluation after candidate search, and subsequently evaluated.
The candidate search around each of the P center points is performed in parallel, within a synchronous Masterslave parallel framework. The master process selects the P center points (Step 2.2) and sends one center point to each slave process. There are P slave processes. Let us assume that the center \(c_i\) is sent to slave process i for a fixed \(i \in \left\{ 1, \dots , P\right\} \). Slave process i performs a surrogateassisted candidate search around the center point \(c_i\) to select a new evaluation point, \(x^{\text {new}}_i\), for expensive evaluation before returning the evaluated point to the master process. The rest of this section provides a description of the mechanism of the surrogateassisted candidate search used to select \(x^{\text {new}}_i\) from a center \(c_i\).
The surrogateassisted candidate search mechanism employed in SOP is referred to as the candidate search method, and is based on the idea of a candidate point method in which we randomly generate a large number of candidate points around the center \(c_i\), and then evaluate them on the surrogate. Müller and Shoemaker [14] compared the candidate point method to the alternative, which is to search for a minimum with an optimization method on the surrogate surface. Based on comparison of a number of test functions and surrogate types, they found no statistical difference between overall algorithm performance when the surrogate surface is searched either with candidate points or with an optimization search such as Genetic Algorithm.
In the candidate search method, the center point \(c_i\) is perturbed to generate a set of \(N_{\mathrm{cand}}\) candidate points, \(\left\{ v_{1},...,\, v_{N_{\mathrm{cand}}}\right\} \).^{4} For every candidate point \(v_{j}\) where \(j = 1, \dots , N_{\mathrm{cand}}\), we randomly select the dimensions of \(c_i\) that will be perturbed, where each coordinate of the center \(c_i\) has a probability of \(\varphi (n)\) to be perturbed. If no dimension k of \(c_{i}\) is selected for perturbation, then one dimension is chosen at random. Any candidate point is then generated by perturbing the selected variables of the center.
Equation 3 is adapted from [20] with n denoting here the number of optimization iterations rather than the number of function evaluations. The rationale for the use of \(\varphi \left( n\right) \) is to reduce the number of dimensions that are perturbed at any iteration. \(\varphi \left( n\right) \) gets smaller with each iteration, gradually reducing the expected number of dimensions perturbed. An intuitive explanation for why this approach is effective is that after many iterations, the solution is possibly quite a good solution so one probably does not want to change the values in all of a large number of dimensions at once. For many blackbox problems the sensitivity of the objective function to different decision variables may vary, and hence perturbing only a subset of decision space may improve search efficiency. The efficiency of this approach has been demonstrated on a related algorithm in [20], but the evidence supporting that efficiency increases by reducing the number of perturbed dimensions is empirical, and is based on the test problems previously studied.
Two candidate generation methodologies, using truncated normal and uniform random perturbation, are employed within SOP and are referred to as nSOP and uSOP, respectively.
 V1

nSOP: \(v_{j}=c_i+Z,\) where \(Z(k)\sim N_{\mathrm{truncated}}\left( 0,\,r_i^{2};\, a(k),\, b(k)\right) ,\) \(a(k)=\text {lb}(k)c_i(k)\) and \(b(k)=\text {ub}(k)c_i(k)\) for \(k\in I_{{\text {perturb}}},\) and Z(k) is zero for \(k\notin I_{{\text {perturb}}}.\) See Online Supplement for a description of \(N_{\mathrm{truncated}}\), the truncated normal distribution.
 V2

uSOP: \(v_{j}(k)=\mathcal {U}\left[ a(k),\, b(k)\right] \) where \(a(k)=\max (\text {lb}(k),\, c_i(k)r_i),\) \(b(k)=\min (c_i(k)+r_i,\,\text {ub}(k))\) for \(k\in I_{{\text {perturb}}},\) and \(v_{j}(k)=c_i(k)\) for \(k\notin I_{{\text {perturb}}}.\) Here, \(\mathcal {U}[u_{1},\, u_{2}]\) is a uniform random variable on \([u_{1},\, u_{2}]\).
The perturbation neighborhood of \(c_i\) in both of the above random candidate generation methods is dependent upon the candidate search neighborhood radius \(r_i\). This radius \(r_i\) is adaptive. Step 2.4 of SOP provides an overview of how the value adapts for each center point.
After \(N_{{\text {cand}}}\) candidate points are generated, the candidate point whose estimated objective function value (from the surrogate model) is smallest is selected as the next function evaluation point \(x^{\text {new}}_i\). Because the surrogate evaluation is inexpensive, this calculation is very fast, even for large \(N_{{\text {cand}}}\). Finally, the slave process evaluates (via expensive simulation) and returns both \(x^{\text {new}}_i\) and its function value \(f\left( x^{\text {new}}_i\right) \) to the master process.
2.4 Step 2.4 adaptive learning and tabu archive
Once new points have been selected and evaluated through simultaneous candidate searches in Step 2.3, Step 2.4 of SOP archives the performance of the search at the end of the current algorithm iteration. This is done for adaptive learning and for subsequently changing the search mechanism of the algorithm.
There are two core components of the adaptive learning process of SOP. In the first component, the algorithm evaluates the candidate search around each center point \(c_i\) as a success or a failure so as to (1) update the candidate search neighborhood radius \(r_i\) of the center point \(c_i\), and (2) update the failure count corresponding to the center point.
Let \(S^{(n)}_{\mathrm{Pareto}}\) be the set of points in the decision space on Front 1 (Pareto set) based on the biobjective problem in Eq. 2 derived from the set \(S^{(n)}\).
If \(x^{\text {new}}_i\) does not generate significant hypervolume improvement, i.e. \(\text {HI}_i\) is smaller than a predefined threshold \(\tau \), then the candidate search around \(c_i\) is tagged as a failure. In this case the algorithm search adapts via reduction of candidate search neighborhood of the center point \(c_i\) by a factor of two. This is done to concentrate the search around closer to this center point, since a wider search radius did not induce an improvement. Furthermore, the failure count of the center point is increased by one. If the failure count goes beyond a predefined threshold, \(N_{f\text {ail}}\), the center point is added to the Tabu List, as it did not produce an improvement in the explorationexploitation tradeoff after multiple candidate searches.^{6}
However, we do not wish to keep a center point in the Tabu List forever. This is because as the algorithm progresses and new points are evaluated, the surrogate model is updated and it may be beneficial to reconsider a tabu point for center selection. Hence, SOP also incorporates a mechanism for removing a point from the Tabu List. In the second component of the adaptive mechanism, the Tabu List is updated, where a center point is removed from the Tabu List if it has been in the Tabu List for the last \(N_{\mathrm{tenure}}\) algorithm iterations. \(N_{\mathrm{tenure}}\) is referred to as the tenure length and is an algorithm parameter.
2.5 Convergence of nSOP
Theorem 1
Suppose that \(x^{*}=\min _{x\in \mathcal {D}}f(x)>\infty \) is the unique global minimizer of f in \(\mathcal {D}\) such that \(\min _{x\in \mathcal {D},\,\left\ xx^{*}\right\ \ge \eta }f(x)>f(x^{*})\) for all \(\eta >0\). If \(\inf _{n\ge 0}\varphi (n)>0,\) nSOP converges almost surely to the global minimum.
The proof of convergence of the nSOP algorithm can be intuitively explained. First, in each iteration, each dimension of the P centers has a boundedawayfromzero probability of being perturbed. Second, the range of sampling for a variable is a truncated normal distribution covering the entire compact hyperrectangle domain \(\mathcal{D}\). Finally, the variance of the normal distribution (perturbation distribution) is bounded above zero because it can only be reduced in half at most \(N_{\mathrm{fail}}\) times. Based on these three conditions, the probability of trial points being inside a ball with a small radius centered at any point x in the compact domain is bounded away from zero. Hence, as the number of iterations goes to infinity, the sequence of random trial points visited by the algorithm is dense in \(\mathcal{D}\). The proof in the Online Supplement addresses the convergence of stochastic nSOP by building on an earlier proof with only one center [18], which is in turn based on Theorem 2.1 in [24] for stochastic algorithms.
This argument that the trial points form a dense set does not apply to the uSOP version. This is because when a dimension is perturbed, the range of perturbation does not cover the whole domain \(\mathcal{D}\), so every ball around a point in the domain does not necessarily have a positive probability of being sampled.
3 Numerical experiments
3.1 Alternative parallel optimization algorithms
We compare our algorithm to Parallel Stochastic RBF [19] and an evolutionary algorithm that uses radial basis function approximation (ESGRBF) [17, 21]. Parallel Stochastic RBF was used to solve an optimization problem arising in groundwater bioremediation introduced in [13], and ESGRBF was used to calibrate computationally expensive watershed models. Regis and Shoemaker reported good results for their Parallel Stochastic RBF method compared to many alternative methods including asynchronous parallel pattern search [8] and a parallel evolutionary algorithm. ESGRBF was shown to significantly outperform conventional Evolution Strategy (ES) algorithms and ES algorithms initialized by SLHDs [17]. We thus compare our algorithms with these two methods.
3.2 Experimental setup
All algorithms are run with \(P=8\) and \(P=32\) function evaluations per iteration. We use the notation \(\mathcal {A}\)8P and \(\mathcal {A}\)32P to distinguish the number of function evaluations per iteration of algorithm \(\mathcal {A}\). For example, Parallel StochRBF algorithm simulating 8 function evaluations per iteration is denoted by StochRBF8P.
We use a cubic RBF interpolation model [16] for the surrogate and Latin hypercube sampling [27] to generate the initial evaluation points for all three examined algorithms. The size of the initial experimental designs was set to \(n_{0}=\min \left\{ n\in \mathbb {N}:\, n\ge 2(d+1)\,\text {and}\, Pn\right\} \), which is the smallest integer larger than \(2(d+1)\) and divisible by P. The number \(2(d+1)\) has previously been used and shown to be an efficient size for initial experimental data set (see e.g. [18, 19]).
The definition of \(n_{0}\) is based on the assumption that (1) each function evaluation takes approximately the same time and (2) P parallel processors are available so that the P expensive evaluations can be distributed to these P processors to use the available computing power efficiently.
Parameter values for SOP
Parameter  Value  Corresponding step 

\(N_{\mathrm{cand}}\)  min (500d, 5000)  Step 2.3 
\(r_{\mathrm{int}}\) (nSOP)  \(0.2\times l\left( \mathcal {D}\right) \)  Steps 2.2–2.4 
\(r_{\mathrm{int}}\) (uSOP)  \(0.1\times l\left( \mathcal {D}\right) \)  Steps 2.2–2.4 
\(N_{\mathrm{fail}}\)  3  Step 2.4 
\(N_{\mathrm{tenure}}\)  5  Step 2.4 
\(\tau \)  \(10^{5}\)  Step 2.4 
3.3 Test functions
We use ten benchmark functions taken from the BBOB test suite [7], namely the functions F15–F24. All of them are 10 dimensional problems and multimodal. The functions are listed in the Online Supplement. For definition and properties of these functions, refer to [7].
3.4 Progress curve in wallclock time
Although the test functions are computationally inexpensive, we assume that each function evaluation takes one hour computation time and that other computational overhead arising, for example, from updating the response surface is negligible. When the objective functions are computationally expensive and function evaluations are done in parallel, the stopping criterion for the optimization is typically a given limit on the allowable wallclock time. Under the assumption that P function evaluations are simulated simultaneously in each unit of wallclock time, we plot a progress curve as a function of wallclock time. The progress curve enables us to compare the performance of the different algorithms over a range of the allowable computation time.
We set the maximum wallclock time to be 60 h. The total number of function evaluations (\(=60P\)) for \(P=8\) and \(P=32\) will then be 480 and 1920, respectively.
3.5 Experimental results and discussion
Figure 5 shows the progress curves of selected test functions. The mean of the best objective function value is plotted on the vertical axis and the wallclock time is plotted on the horizontal axis. The figure shows both the case for doing simultaneously 8 expensive evaluations and 32 parallel evaluations for a comparison. For each test function, in addition to the main plot, the last 15 iterations are plotted separately in the small subfigure so that the tail of the plot before the algorithm terminates can be seen clearly.
Overall, SOP together with either normal or uniform strategies leads to a very good performance, clearly outperforming the other two alternative methods. We also find that for \(P=8,\) StochRBF8P performs better than ESGRBF8P. However, when \(P=32,\) ESGRBF32P performs better than StochRBF32P.
When doing more function evaluations per iteration, it should be expected that the algorithm improves the objective function value in less wallclock time since more information is obtained for refining the response surface. However, in some cases, SOP8P outperformed the alternative methods that do 32 evaluations per iteration. For example, for function F15, nSOP8P and uSOP8P with 8 processors got better results in a shorter wallclock time than StochRBF32P and ESGRBF32P. For function F16, both ESGRBF8P and ESGRBF32P are worse than nSOP8P, uSOP8P, and StochRBF8P. In addition, ESGRBF8P even outperformed ESGRBF32P after around 15 h indicating that parallel ESGRBF is not efficient for higher number of processors on some functions.
As for functions F21 and F22, nSOP and uSOP converge fastest and to a better final solution. Moreover, nSOP8P and uSOP8P again outperformed both StochRBF32P and ESGRBF32P by finding a lower objective function value in less wallclock time. StochRBF32P did very poorly on these two test functions and StochRBF8P even surpassed StochRBF32P in both functions.
In contrast to Stochastic RBF and ESGRBF, SOP with 32 processors always found the solution with lower value in less wallclock time than for SOP with 8 processors, indicating that SOP was much more efficient at utilizing higher number of processors.
Additional results can be found in the Online Supplement. In particular, there is a comparison of the number of problems on which nSOP or uSOP is significantly better than alternative algorithms in terms of the final solution. In this regard, both nSOP and uSOP are significantly better than Parallel StochRBF on three and six (while worse on one and no) problems when using 8 and 32 processors, respectively. We note that these results in the Online Supplement are based only on the final values at the maximum number of iterations. However, looking at the progress curves (Fig. 5), the reader can see that in many cases SOP could achieve accurate solutions much more quickly (e.g. less than 20 h of wallclock time) than the other methods, which is not reflected in these tables.
In addition to the BBOB testbed, we also tested our algorithms on a 12dimensional problem arising in the detoxification of contaminated groundwater using aerobic bioremediation [28]. The description of this problem is given in the Online Supplement. The parameters \(\mu =26,\) \(\lambda =160,\) and \(\nu =64\) are set for ESGRBF64P. All other algorithm parameters are the same as those used for BBOB testbed, and the results are given in Fig. 6.
Results for GWB12D after 60 h
Alg  nSOP  uSOP  StochRBF  ESGRBF  

P  Mean  Std  Mean  Std  Mean  Std  Mean  Std 
(a) Mean and standard deviation of the best function value after 60 h for \(P=8\), 32 and 64 processors on GWB12D. The best value for each P is shown in bold font  
8  350.964  17.940  346.727  15.606  390.614  31.145  377.373  27.752 
32  329.166  8.034  326.112  5.016  342.550  13.997  329.128  9.875 
64  318.774  2.465  319.182  3.658  332.473  4.390  325.420  8.837 
P  Mann–Whitney–Wilcoxon test results  

(b) Mann–Whitney–Wilcoxon test results for GWB12D at 5 % significance level  
8  nSOP \(\approx \) uSOP \(\prec \) ESGRBF \(\approx \) StochRBF  
32  nSOP \(\approx \) uSOP \(\approx \) ESGRBF \(\prec \) StochRBF  
64  nSOP \(\approx \) uSOP \(\prec \) ESGRBF \(\prec \) StochRBF 
3.6 Relative speedup
The relative \(\alpha {\text {Speedup(P)}}\) for each test function is calculated for different \(\alpha \)levels, \(\alpha _{1}>\alpha _{2}>\alpha _{3}\) of accuracy. To get the three \(\alpha \)levels of nSOP for test functions in BBOB testbed, nSOP(8) and Stochastic RBF are run for 496 function evaluations and nSOP(32) for 1984 function evaluations. Let \(y_{1}^{*},\) \(y_{8}^{*},\) and \(y_{32}^{*}\) be the average best objective function values obtained from Stochastic RBF, nSOP(8) and nSOP(32), respectively. We set \(\alpha _{3}=\max \left\{ y_{1}^{*},\, y_{8}^{*},\, y_{32}^{*}\right\} \), \(\alpha _{2}=\alpha _{3}+\left \alpha _{3}\right \times 0.01\) and \(\alpha _{1}=\alpha _{3}+\left \alpha _{3}\right \times 0.05\). For the GWB12D test function, we ran in addition nSOP(64) for 3968 function evaluations. Let \(y_{64}^{*}\) be the average best objective function values obtained from nSOP(64), and define \(\alpha _{3}=\max \left\{ y_{1}^{*},\, y_{8}^{*},\, y_{32}^{*},\, y_{64}^{*}\right\} \), \(\alpha _{2}=\alpha _{3}+\left \alpha _{3}\right \times 0.01\) and \(\alpha _{1}=\alpha _{3}+\left \alpha _{3}\right \times 0.05\).
\(\alpha \)Speedup of nSOP and uSOP
Test function  P  nSOP  uSOP  

\(\alpha _{1}\)  \(\alpha _{2}\)  \(\alpha _{3}\)  \(\alpha _{1}\)  \(\alpha _{2}\)  \(\alpha _{3}\)  
F15  8  11. 805  11.976  12.195  11.524  11.690  11.905 
32  19.360  19.640  20.000  20.167  20.458  20.833  
F16  8  2.318  2.545  3.159  3.105  3.182  4.889 
32  5.100  4.308  5.378  6.556  5.600  9.059  
F17  8  10.333  23.706  27.500  8.857  21.211  24.750 
32  24.800  57.571  61.875  20.667  44.778  49.500  
F18  8  17.529  23.000  24.800  12.417  15.607  17.103 
32  37.250  48.556  55.111  29.800  39.727  45.091  
F19  8  8.800  6.500  5.689  7.600  9.639  8.267 
32  29.333  37.143  34.700  22.800  31.545  19.840  
F20  8  5.636  5.333  2.032  6.200  6.400  2.143 
32  20.667  16.000  3.048  15.500  12.800  3.649  
F21  8  3.818  14.056  19.192  3.500  14.056  10.848 
32  8.400  31.625  38.385  7.000  31.625  38.385  
F22  8  12.391  10.258  10.578  17.375  13.714  4.698 
32  40.714  31.800  8.207  39.714  36.000  32.889  
F23  8  2.000  8.375  1.746  2.000  13.800  2.651 
32  2.000  33.500  5.000  2.000  23.000  4.071  
F24  8  7.275  9.226  8.197  6.577  7.220  7.267 
32  13.741  14.382  14.706  24.429  20.286  18.957  
GWB12D  8  3.850  3.255  3.000  4.289  3.345  3.397 
32  9.059  8.950  9.000  8.150  7.462  7.926  
64  12.833  11.933  11.812  12.538  12.125  12.588 
The algorithm results will change as P changes since P centers are used to generate candidate points and the response surface is updated only once every P evaluations. On some problems, the change is quite helpful and actually reduces the number of evaluations required to reach an \(\alpha \)level. In this case, the speedup is “superlinear”, i.e. \(\alpha {\text {Speedup(P)}}\) is greater than P. The superlinear speedup holds at many \(\alpha \)levels for problems F17, F18, F21, and F22. F17 is Schaffers function, which is highly multimodal– both the frequency and amplitude of the modulation vary. F18 is a moderately illconditioned counterpart to F17, with conditioning of about 1000. Both F21 and F22 are Gallagher’s Gaussian function, where F21 consists of 101 local optima, and the conditioning around the global optimum is about 30. F22 consists of 21 local optima and the conditioning around the global optimum is about 1000 [7].
The three functions that have rather poor scalability (i.e. speedup low compared to P) are F16, F20, and F23. These functions are described in [7] as being highly rugged and moderately to highly repetitive.
For F15, the superlinear speedup holds for 8 processors but not for 32 processors. As for F19, and F24, although they do not achieve a superlinear speedup, we can see that the speedup seems to improve in the number of processors. These three functions, despite being highly multimodal, are not as rugged as those mentioned in the previous paragraph. The global amplitude of F15 is large compared to local amplitudes. As for F24, the function was constructed to be deceptive for some evolutionary algorithms with large population size. This might be the reason why the scalability of SOP on this test function deteriorates with 32 processors.
3.7 Sensitivity of SOP to \(N_{\mathrm{fail}}\)
A larger \(N_{\mathrm{fail}}\) allows the algorithm to search longer around a particular center with no improvement until it is declared as tabu (search more locally). On the other hand, a small \(N_{\mathrm{fail}}\) allows the algorithm to leave the region around centers with no improvement earlier (search more globally).
To examine the effect of changing \(N_{\mathrm{fail}}\), SOP is implemented also with \(N_{\mathrm{fail }}=0\) (which is the extreme case where any unimproved center is immediately tabu) and \(N_{\mathrm{fail}}=10\) (longer length before being tabu), while our default implementation is \(N_{\mathrm{fail }}=3\). Figure 7 shows the comparison between these implementations on some problems with nSOP8P and uSOP8P.
The results suggest that the best \(N_{\mathrm{fail}}\) varies from problem to problem. For example, for uSOP8P (Fig. 7b), on the test function F18, the algorithm implemented with \(N_{\mathrm{fail}}=10\) results in faster progress while the algorithm with the default \(N_{\mathrm{fail}}=3\) is better on the test function F15. In addition, for the same test problem, the best \(N_{\mathrm{fail}}\) of nSOP and uSOP can also be different (e.g. F18).
We briefly discuss the results using a large number of processors without providing the plots. When \(P=32\), similar results are obtained when using \(N_{\mathrm{fail}}=3\) and \(N_{\mathrm{fail }}=10\). This is because when P is large, multiple tasks take place at the same time around each of the P centers, and so the memory archive for tabu list does not matter that much. Also, we found that in one case the result obtained with \(N_{\mathrm{fail }}=0\) is best. Although this is not typical, the reason might be that immediate tabu allows the algorithm to explore the surface more globally.
4 Conclusions
Parallel computation has the potential to greatly reduce the wallclock time to solve a global optimization problem for a computationally expensive objective function, but that potential can only be realized if the parallel algorithm is able to effectively select the work to be computed in parallel. The efficiency of the parallel surrogate algorithms depends on the P decision vectors selected for function evaluations in each iteration to provide the most useful information over the course of many iterations. We want these values to both help improve the surrogate surface accuracy and to help identify the neighborhood of the global minimum.
In this paper we introduced the algorithm, SOP, which is based on the selection of center points through nondominated sorting. To improve the diversity, promising points on the sorted fronts whose function values are small and are far away from other evaluated points are selected as centers. The selected centers are then used for generating a set of candidate points from which the next function evaluation points are chosen. Multiple centers generate a more diverse set of candidate points. We also incorporate a tabu structure to further diversify the search by not allowing some points on the better fronts to be selected under certain conditions. Two versions of SOP, nSOP and uSOP, relying on the different candidate generation methodologies are proposed. It can be shown that the nSOP algorithm converges to the global optimum almost surely.
In the numerical experiments, we compared SOP with two parallel RBFbased algorithms, Parallel StochRBF and ESGRBF. Our numerical results indicate that overall SOP was more effective than the alternative methods on the synthetic test functions. In some cases, SOP with just 8 processors could obtain a better result in less wallclock time than the alternative algorithms with 32 processors. The results on a groundwater bioremediation problem using up to 64 processors also showed that our algorithm was significantly better than the alternative methods. Hence, these results are expected to have a positive impact in the field of expensive blackbox optimization, which applies to many areas, in particular in (but not limited to) engineering.
Footnotes
 1.
Pseudocode for SOP algorithm is given in Algorithm 1 of the Online Supplement.
 2.
Pseudocode describing the twotier ranking system can be found in Figure A.2 (two_tier_ranking) of the Online Supplement.
 3.
Pseudocode describing the P center selection procedure can be found in Figure A.3 (P_centers_sel) of the Online Supplement.
 4.
Pseudocode describing the candidate search method can be found in Figure A.4 (candidate_search) of the Online Supplement.
 5.
Pseudocode for hypervolume improvement can be found in Figure A.5 (hypervol_improv_indc) of the Online Supplement.
 6.
Pseudocode for tabu archiving can be found in Figure A.6 (update_tabu_archive) of the Online Supplement.
Notes
Acknowledgments
This research was conducted primarily when Dr. Krityakierne was a PhD student in Applied Mathematics at Cornell University. During that time, her time and that of Prof. Shoemaker were supported in part by NSF grants 1116298 (CISE) and 1049033 and DOESciDAC DESC0006791 and by the Fulbright and the King Anandamahidol Foundation of Thailand fellowships to Dr. Krityakierne. All the authors continued to work on the manuscript in their new positions after leaving Cornell. Financial support for completing the manuscript was provided by National University of Singapore.
Supplementary material
References
 1.Akhtar, T., Shoemaker, C.A.: Multi objective optimization of computationally expensive multimodal functions with RBF surrogates and multirule selection. J. Glob. Optim. 64, 17–32 (2015)MathSciNetCrossRefMATHGoogle Scholar
 2.Bischl, B., Wessing, S., Bauer, N., Friedrichs, K., Weihs, C.: Moimbo: multiobjective infill for parallel modelbased optimization. In: Learning and Intelligent Optimization, pp. 173–186. Springer, New York (2014)Google Scholar
 3.Booker, A.J., Dennis Jr, J.E., Frank, P.D., Serafini, D.B., Torczon, V., Trosset, M.W.: A rigorous framework for optimization of expensive functions by surrogates. Struct. Optim. 17(1), 1–13 (1999)CrossRefGoogle Scholar
 4.Deb, K., Agrawal, S., Pratap, A., Meyarivan, T.: A fast elitist nondominated sorting genetic algorithm for multiobjective optimization: NSGAII. Lect Notes Comput Sci 1917, 849–858 (2000)Google Scholar
 5.Deb, K.: MultiObjective Optimization Using Evolutionary Algorithms, vol. 16. Wiley, London (2001)MATHGoogle Scholar
 6.Gutmann, H.M.: A radial basis function method for global optimization. J. Glob. Optim. 19, 201–227 (2001)MathSciNetCrossRefMATHGoogle Scholar
 7.Hansen, N., Finck, S., Ros, R., Auger, A., et al: Realparameter blackbox optimization benchmarking 2009: Noiseless functions definitions (2009)Google Scholar
 8.Hough, P.D., Kolda, T.G., Torczon, V.J.: Asynchronous parallel pattern search for nonlinear optimization. SIAM J. Sci. Comput. 23(1), 134–156 (2001)MathSciNetCrossRefMATHGoogle Scholar
 9.Huang, D., Allen, T.T., Notz, W.I., Zeng, N.: Global optimization of stochastic blackbox systems via sequential kriging metamodels. J. Glob. Optimiz. 34(3), 441–466 (2006)MathSciNetCrossRefMATHGoogle Scholar
 10.Jones, D.R., Schonlau, M., Welch, W.J.: Efficient global optimization of expensive blackbox functions. J. Glob. Optim. 13(4), 455–492 (1998)MathSciNetCrossRefMATHGoogle Scholar
 11.Li, Y., Liu, L., Long, T., Chen, X.: Multipleoptima search method based on a metamodel and mathematical morphology. Eng. Optim. 48(3), 437–453 (2016)MathSciNetCrossRefGoogle Scholar
 12.Liu, H., Shengli, X., Ma, Y., Wang, X.: Global optimization of expensive black box functions using potential lipschitz constants and response surfaces. J. Glob. Optimiz. 63(2), 229–251 (2015)MathSciNetCrossRefMATHGoogle Scholar
 13.Mugunthan, P., Shoemaker, C.A., Regis, R.G.: Comparison of function approximation, heuristic, and derivativebased methods for automatic calibration of computationally expensive groundwater bioremediation models. Water Resour. Res. 41(11) (2005). doi: 10.1029/2005WR004134
 14.Müller, J., Shoemaker, C.A.: Influence of ensemble surrogate models and sampling strategy on the solution quality of algorithms for computationally expensive blackbox global optimization problems. J. Glob. Optimiz. 60(2), 123–144 (2014)MathSciNetCrossRefMATHGoogle Scholar
 15.Paulavičius, R., Sergeyev, Y.D., Kvasov, D.E., Žilinskas, J.: Globallybiased disimpl algorithm for expensive global optimization. J. Glob. Optimiz. 59(2–3), 545–567 (2014)MathSciNetCrossRefMATHGoogle Scholar
 16.Powell, M.J.D.: The theory of radial basis function approximation in 1990. In: Advances in Numerical Analysis, vol. 2: Wavelets, Subdivision Algorithms and Radial Basis Functions. Oxford University Press, Oxford, pp. 105–210 (1992)Google Scholar
 17.Regis, R.G., Shoemaker, C.A.: Local function approximation in evolutionary algorithms for the optimization of costly functions. IEEE Trans. Evol. Comput. 8(5), 490–505 (2004)CrossRefGoogle Scholar
 18.Regis, R.G., Shoemaker, C.A.: A stochastic radial basis function method for the global optimization of expensive functions. INFORMS J. Comput. 19(4), 497–509 (2007)MathSciNetCrossRefMATHGoogle Scholar
 19.Regis, R.G., Shoemaker, C.A.: Parallel stochastic global optimization using radial basis functions. INFORMS J. Comput. 21(3), 411–426 (2009)MathSciNetCrossRefMATHGoogle Scholar
 20.Regis, R.G., Shoemaker, C.A.: Combining radial basis function surrogates and dynamic coordinate search in highdimensional expensive blackbox optimization. Eng. Optim. 45(5), 529–555 (2013)MathSciNetCrossRefGoogle Scholar
 21.Shoemaker, C.A., Regis, R.G., Fleming, R.C.: Watershed calibration using multistart local optimization and evolutionary optimization with radial basis function approximation. Hydrol. Sci. J. 52(3), 450–465 (2007)CrossRefGoogle Scholar
 22.Sobester, A., Leary, S.J., Keane, A.J.: A parallel updating scheme for approximating and optimizing high fidelity computer simulations. Struct. Multidiscip. Optim. 27(5), 371–383 (2004)CrossRefGoogle Scholar
 23.Sóbester, A., Leary, S.J., Keane, A.J.: On the design of optimization strategies based on global response surface approximation models. J. Glob. Optimiz. 33(1), 31–59 (2005)MathSciNetCrossRefMATHGoogle Scholar
 24.Spall, J.C.: Introduction to Stochastic Search and Optimization: Estimation, Simulation, and Control, vol. 65. Wiley, London (2005)MATHGoogle Scholar
 25.Viana, F.A.C., Haftka, R.T., Watson, L.T.: Efficient global optimization algorithm assisted by multiple surrogate techniques. J. Glob. Optimiz. 56(2), 669–689 (2013)CrossRefMATHGoogle Scholar
 26.Vrugt, J. A., Robinson, B. A.: Improved evolutionary optimization from genetically adaptive multimethod search, Proc Nat Acad Sci, 104(3), 708–711 (2007)Google Scholar
 27.Ye, K.Q., Li, W., Sudjianto, A.: Algorithmic construction of optimal symmetric latin hypercube designs. J. Stat. Plan. Inference 90(1), 145–159 (2000)MathSciNetCrossRefMATHGoogle Scholar
 28.Yoon, J.H., Shoemaker, C.A.: Comparison of optimization methods for groundwater bioremediation. J. Water Resour. Plan. Manag. 125(1), 54–63 (1999)CrossRefGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.