Generally, the inconvenience of establishing the mathematical optimization models directly and the conflicts of preventing simultaneous optimization among several objectives lead to the difficulty of obtaining the optimal solution of a practical engineering problem with several objectives. So in this paper, a generate-first-choose-later method is proposed to solve the multiobjective engineering optimization problems, which can set the number of Pareto solutions and optimize repeatedly until the satisfactory results are obtained. Based on Frisch’s method, Newton method, and weighed sum method, an efficient hybrid algorithm for multiobjective optimization models with upper and lower bounds and inequality constraints has been proposed, which is especially suitable for the practical engineering problems based on surrogate models. The generate-first-choose-later method with this hybrid algorithm can calculate the Pareto optimal set, show the Pareto front, and provide multiple designs for multiobjective engineering problems fast and accurately. Numerical examples demonstrate the effectiveness and high efficiency of the hybrid algorithm. In order to prove that the generate-first-choose-later method is rapid and suitable for solving practical engineering problems, an optimization problem for crash box of vehicle has been handled well.
Most of the practical engineering optimization problems are multiobjective. For example, an airplane design problem might require maximizing fuel efficiency and payload, while minimizing the weight of the structure [
Recently, researchers have proposed various methods with the fast development of multiobjective optimization. Generally, these methods can be divided into scalar methods and evolution methods by way of solving the optimization problems.
Scalar methods transform the vector optimization problems into adaptive scalar ones. Combining with some gradient methods, a scalar method can easily obtain the optimum by iterations. Typical multiobjective scalar methods include traditional weighted sum method [
The idea of evolutionary methods is similar to the biological evolution process, established by Darwin’s theory of natural selection. Referring to the biological evolution process, evolutionary methods make use of crossover, mutation, or inheritance operations during iterations to obtain better results. Evolutionary methods have developed rapidly in the field of multiobjective optimization, such as Multiobjective Genetic Algorithm [
In recent years, the multiobjective optimization methods are widely used in the engineering problems, such as aerospace [
The purpose of the method proposed in this paper is to provide the design advice to designers quickly. So we hope that, during the process of calculation, the method can rapidly get the Pareto optimal solutions and Pareto front of the multiobjective optimization problems. Meanwhile, if the results are not satisfactory, it is necessary to increase the number of the solutions or reconstruct the models of multiobjective optimization problems. Doubtlessly, the low calculation efficiency of multiobjective evolutionary algorithms dissatisfies this need.
Usually, the engineering optimization problems cannot be solved directly. They can be described approximately by surrogate models and then can be optimized. The surrogate models are always constructed by polynomial response surface method [
We should consider that the surrogate models are meaningful only in the constraint interval. So in order to keep the computational accuracy, this paper takes Frisch’s method [
For obtaining excellent crashworthiness performance of vehicles, some researchers studied about the structural optimization problems. Acara et al. took CFE and SEA as optimization objectives and got a tradeoff solution by sequential quadratic programming [
The outline of this paper is as follows. Section
Generally, in engineering optimization problems there are many objectives which are always conflicting with each other. Because of the complex conditions and structural shape, the relationship among objective functions, design variables, and constraint functions is hard to construct directly. Sometimes, in practical engineering problems, numerical simulation and experiments are adopted, but these ways rely on the experience of designers and make it hard to achieve the optimal design globally.
Surrogate models are widely used to improve the efficiency of engineering design and optimization. In this paper, a valid method for solving the engineering optimization problems based on surrogate model is proposed, which is a generate-first-choose-later method. After constructing the surrogate models of multiobjective engineering optimization problems with constraints, the proposed method can calculate the Pareto optimal solutions and Pareto front. The Pareto front is shown intuitively to provide lots of suggestions for designers as a reference. Meanwhile, designers can reset the number of solving Pareto optimal solutions and calculate again, in order to get better results. The flow chart of the proposed method is shown in Figure
The generate-first-choose-later method for multiobjective optimization engineering problems.
In the proposed method, constructing the surrogate models of engineering problems is very important, which will highly influence the accuracy of optimization results. After determining the optimization problem, the surrogate models of objective functions and constraint functions can be constructed by response surface method, radial basis function method, Kriging method, and so on.
It is worth noting that the magnitudes of different objective functions are often different. So we have to unify the expression and normalize the functions before optimization. The theory of the proposed method will be stated in Section
Evolutionary methods are widely used in engineering multiobjective optimization problems recently. Moreover, the derivative of some engineering problems does not exist or cannot be obtained easily. So the evolutionary methods are more suitable for solving these problems. However, the disadvantages of evolutionary methods should not be neglected [
The randomness of searching in the iterative directions leads to the slow convergence rate of evolutionary methods. When the individuals are far from the Pareto optimal solutions, both the field which can be Pareto improved and the opportunity to generate a descent direction randomly are large. With the approach to Pareto optimal solutions, the conflict among objective functions increases, which leads to the difficulty of finding the descent directions for each objective function. When the individuals are close to the Pareto optimal solutions, both the proportion of the fields which can be Pareto improved and the opportunity to generate a descent direction randomly are small. These are the reasons that convergence rate of evolutionary methods solving the multiobjective optimization problems is fast during the initial stage and slow during the final stage.
Compared with multiobjective evolutionary methods, the advantage of local search methods in efficiency is remarkable, such as Newton method.
The models of multiobjective engineering optimization problems in this paper are established by response surface methods, and also the constraints are handled by log functions. So the optimization models are derivable and the second derivative can be obtained by gradient methods. Meanwhile, the method proposed in this paper may calculate Pareto optimal solutions more than once in order to provide satisfying advice for designers. Hence, the computational efficiency of the method is very important.
Although evolutionary methods are widely used in multiobjective engineering optimization problems, the computational efficiency is not satisfactory. In this paper, a multiobjective scalar method is researched, which has the advantage of fast convergence. Newton method has been chosen to calculate Pareto optimal solutions in this paper, for its high computational efficiency.
The iteration direction of Newton method includes the gradient and Hessian matrix information of objective functions, so the iteration point can be definitely close to the optimal point. When the iteration point is near the optimal point, the rate of convergence is rapid [
Newton method is chosen as the main calculation algorithm for searching the solutions. The process of solving multiobjective engineering optimization problems with upper and lower bounds is described in detail. And the overall construction course of the hybrid algorithm is in the following.
In practical engineering optimization problem, there are always several objects which are conflicting to prevent simultaneous optimization of each other. There is no one optimal solution satisfying all the minima of objects. So searching for the Pareto optimal set of these objects is one of the most effective ways. Meanwhile, there are some constraints in the engineering optimization problems, generally about the upper and lower bounds of design variables. In this section, a valid method will be proposed, in order to fast calculate the multiobjective optimization problem with upper and lower bounds.
The
For convenient calculation, the upper and lower bounds can be transformed into upper and lower bounds; that is,
So the multiobjective optimization engineering problem with
The aim of solving multiobjective optimization problems is to obtain the Pareto optimal set. For two design variables
The process of solving multiobjective optimization problems with constraints is to find the Pareto optimal set of all the objects in the feasible region under the constraint conditions. The problem should be transformed into an unconstrained one first and then be solved. Hence, a penalty term can be added to the object functions. When the penalty term is closer to zero, it means the design variables satisfy the constraints. During the solving process, the penalty term should be scaled down until it is small enough and can be neglected relative to the object values for meeting the stopping criterion. At the moment, the obtained solution can not only be equivalent to the optimal solution of the original problem but also satisfy the constraints.
The optimization problems discussed in this paper are based on surrogate models. So the interior point method is chosen to deal with the constraints. For solving the gradient and Hessian matrix conveniently, the penalty term is constructed by Frisch’s method, which is expressed as
Despite having deficiencies in depicting the Pareto optimal set, the weighted sum method for multiobjective optimization continues to be used extensively not only to provide multiple solution points by varying the weights consistently but also to provide the solutions that reflect the preference of each object. In the proposed method of this paper, weighted sum method is chosen. After the designers set the number of Pareto optimal solutions, the corresponding group of weighting factors can be provided uniformly.
Because the gradient and Hessian matrices of object and constraint functions can be obtained, Newton method is selected for its rapid convergence. In order to improve computational efficiency, the negative gradient direction is considered when the Newton directions are not descent. In this paper, the Pareto optimal solutions will be solved by iteration. The process of deducing iteration direction is as follows.
First, the penalty functions should be constructed. The logarithmic penalty function of the
Express
For
Then, adding the penalty function of every object function together, a sum function can be expressed as
By calculating the derivative of
This method combines Newton method and linear weighted sum method, and the iteration direction
In order to ensure all the iteration directions are descending during the optimization, an identification process is introduced. If Newton direction at some point is ascending, take the negative gradient direction of sum function at this point as the iteration direction. The criterion is the product of Newton direction and negative gradient direction, denoted as
So the iteration direction at
During the process of calculation, the selection of iteration step length is necessary. For the object functions are based on surrogate models, the accuracy can be ensured only if the design variables satisfy the constraint conditions. Hence, a criterion is set to prevent that the design variables dissatisfy the constraint conditions. When the iteration point is beyond the range of constraints, the step length will be scaled down. Until the new iteration point satisfies the constraints, output the current step length.
In this section, details of calculating Pareto optimal solutions then forming the Pareto optimal set and Pareto front will be described.
As in the introduction above, proposing a method to obtain Pareto optimal solutions rapidly is the key of this paper. In this paper, based on Frisch’s method, Newton method, and weighted sum method, an effective algorithm is put forward, named Algorithm
The whole process of Algorithm
Establish the logarithmic penalty functions of each objective
Choose an initial point
Calculate the gradients of each logarithmic penalty function
If
Else, go to Step
Calculate the iterative direction
If
Else,
Calculate iteration step size
If for all of the constraint functions
Else, go to Step
Consider
Iteratively calculate
Calculate penalty term
Else,
In the process, the small positive constants
Algorithm
The flow chart of the hybrid algorithm for constrained multiobjective optimization problems.
According to this method, designers should only set the number of solutions and initial point. By calculating automatically, a Pareto optimal set will be output. In addition, the Pareto front will be shown as a coordinate graph for designers.
The solution of multiobjective function optimization problems obtained by the present method is always a local solution, which converges to the real Pareto front when the stopping criterion is close to zero. If the objective functions are convex then the local solution is global one at the same time. By the popular evolutionary algorithms, global solution can be got, which turns out to be not close to the real Pareto front in a short time. When the constrained and objective functions are continuously differentiable and nonlinear, the solution close to the real Pareto front can be got rapidly by the proposed method. So, a more accurate solution can be obtained by the proposed algorithm in a short time. However, a good solution is hard to obtain by the algorithm when the constrained and objective functions are not continuously differentiable and nonlinear.
Another advantage by the present method is high efficiency in converging to the Pareto front, and the shortcoming is that the objective functions and inequality constraints must be continuously differentiable and nonlinear [
On one side, many multiobjective engineering optimization problems can be established as the mathematical models which are nonlinear and continuously differentiable, such as the engineering problems described by surrogate models. On the other side, the hybrid method provides references to designers more rapidly than popular evolutionary algorithms, which will improve working efficiency apparently. So, the method proposed in this paper presents an applicable value for actual engineering optimization problems.
The solutions by different algorithms are compared in detail next.
Two benchmark numerical examples are chosen to check the algorithm. One of them is an example in the user’s guide of MATLAB [
For further evaluation, the tests are also executed by Multiobjective Genetic Algorithm. Then, the results obtained by the proposed method are compared with the one by Multiobjective Genetic Algorithm. The methods’ performances are evaluated from three aspects, which include the diversity of solutions in Pareto front, the accuracy of the Pareto solutions, and the computational efficiency. To keep things simple, the Multiobjective Genetic Algorithm is written as MOGA for short and the proposed algorithm is named as NSWFA. In all tests, one hundred initial points are iterated for a Pareto optimal set. The Pareto optimal fronts of the two tests are shown in Figures
The Pareto optimal front of test 1 obtained by the two algorithms.
The Pareto optimal front of test 2 obtained by the two algorithms.
Sometimes, the solutions of multiobjective optimization problems, which are obtained by the algorithms based on weighted sum method, are not well distributed in the Pareto optimal front. Studies can be found in this field [
The solutions obtained by NSWFA are closer to real Pareto optimal front than MOGA.
The solutions obtained by NSWFA are a little closer to real Pareto optimal front than MOGA.
Another important performance is the computational efficiency, which is studied from the iterative number and the consumed CPU time. Facts have proved that the iterative points cannot be convergent to the Pareto optimal front in a short time by evolutionary algorithms.
In this paper, the maximum number of iterations is 2000, and the results by MOGA at the 2000th iteration are recorded. More detailed information is listed in Table
The detailed comparison of the two algorithms in computation efficiency.
Algorithm | Stopping criteria | Initial points | Obtained points | CPU time | Iterations | Convergence | |
---|---|---|---|---|---|---|---|
Test 1 | MOGA | 10−5 | 100 | 70 | 329.72 s | 2000 | No |
NSWFA | 10−5 | 100 | 100 | 13.23 s | 730 | Yes | |
|
|||||||
Test 2 | MOGA | 10−5 | 100 | 77 | 361.88 s | 2000 | No |
NSWFA | 10−5 | 100 | 100 | 5.25 s | 615 | Yes |
The stopping criterion is 10−5 and the results by MOGA are all not convergent. But the results by NSWFA are convergent to the Pareto optimal front with less iterations and CPU time. At each iteration, the whole population is iterated at the same time by MOGA, but only one point is iterated by NSWFA. The average iterations of one Pareto optimal solution are only 7.3 and 6.15 times by NSWFA. That is why the CPU time by NSWFA is far less than MOGA. In addition, more Pareto optimal solutions are got by NSWFA than MOGA with the same initial points.
Generally, the Pareto optimal solutions obtained by NSWFA have better accuracy and spread than MOGA. Also, far less CPU time is consumed by NSWFA than MOGA. All the evidence suggests that NSWFA is perfect in the multiobjective optimization problems whose mathematical models and inequality constraints are nonlinear and continuous.
The crash box is an important part of car collision system, which plays an important role in occupant protection during the collision of vehicles at low speed. In the section, the generate-first-choose-later method with the hybrid algorithm proposed for multiobjective engineering optimization design problems in this paper is adopted to finish the design of the crash box.
The properties of energy absorption and maximum crushing force must be considered simultaneously in designing the crash box. The structure design becomes a problem of complicated multiobjective optimization design. The crash box is made of four cutting boards whose thickness can be chosen as design variables to optimize. In the collision of vehicles at low speed, the crash box should absorb the collision energy as much as possible, but the peak force should be small as soon as possible. So in this problem, energy absorbing and the biggest impact are selected to be objectives, and the four wall thickness values of the crash box are selected as design variables. A car collision system with two crash boxes is shown in Figure
The location of crash box.
The 4 design variables of crash box.
In order to effectively simulate the energy absorption characteristic of crash box under axial load in the vehicle frontal impact, choose the bumper and crash box as a whole for research according to real crash process.
The rear part of crash box connecting vehicle body is constrained, while the front part is struck by a rigid wall weighing one ton with speed of 4 m/s. In the simulation of low-speed crash, the model could use elastoplastic material without regard to strain-rate effect. The response surface models are established referring to the paper of Li et al. [
Adopting quadratic regression orthogonal combination design of experiment and distributing experiment point reasonably, 25 simulation experiments are completed.
According to the result, the response surface models of energy absorption
After getting the surface functions, the variance analysis is used to verify the fitting degree. In the process, the determination coefficient
In the formula,
In general, if the determination coefficient and adjusting determination coefficient are more close to 1, the response surface function with respect to response variables is more precise. The determination coefficients
In order to reduce passenger injury, the crash box is desired to absorb more energy and generate less maximum crushing force, that is, maximize the objective function
The standard multiobjective optimization model should be established firstly. When converting the objective function to minimization problem, the magnitude of objective functions should be comparable since large magnitude results in large deviation. Based on simulation result, the minimum energy absorption (3365.7 J) and maximum crushing force (144.129 kN) are chosen to build corresponding objective function. The constraint of crash box design is upper and lower limitation of wall thickness. The blanking plate should be between 1 mm and 3 mm due to processing factor, that is, wall thickness
To solve constrained multiobjective optimization problem by the proposed method, set 100 as the number of Pareto optimal solutions and pick a random initial design variable in accordance with constraint. The Pareto optimal front of standard model by rapid calculation is shown in Figure
The Pareto optimal front of the standard multiobjective optimization model.
According to Pareto optimal solution set, the corresponding energy absorption and maximum crushing force are obtained to form Pareto front of real crash problem, which helps engineers choose design proposal more conveniently. The Pareto front end of crash problem is shown in Figure
The corresponding properties of the Pareto optimal solutions.
The design of crash box mainly aims to improve performance. However, the mass of structure is also a very important factor. The mass of crash box varies according to the design variables. Their relation is geometrical. It is easy to get mass based on design variables in the simulation software. The coordinates graph about the energy absorption, maximum crashing force, and mass is shown in Figure
The referable properties and the corresponding mass.
Figure
The results of 100 pieces of design advice for reference.
|
|
|
|
|
|
||||
---|---|---|---|---|---|---|---|---|---|
|
|
|
|
|
| ||||
1 | 0.01 | 0.99 | 3.0000 | 1.0000 | 1.0000 | 1.0000 | 2699.0 | 40.0865 | 2.4183 |
2 | 0.02 | 0.98 | 3.0000 | 1.0000 | 1.0000 | 1.0000 | 2699.0 | 40.0865 | 2.4183 |
3 | 0.03 | 0.97 | 3.0000 | 1.0000 | 1.0000 | 1.0000 | 2699.0 | 40.0865 | 2.4183 |
4 | 0.04 | 0.96 | 3.0000 | 1.0000 | 1.0000 | 1.0362 | 2737.0 | 40.1824 | 2.4250 |
5 | 0.05 | 0.95 | 3.0000 | 1.0000 | 1.0000 | 1.0937 | 2797.5 | 40.3628 | 2.4357 |
6 | 0.06 | 0.94 | 3.0000 | 1.0000 | 1.0000 | 1.1486 | 2855.4 | 40.5669 | 2.4458 |
7 | 0.07 | 0.93 | 3.0000 | 1.0000 | 1.0000 | 1.2011 | 2910.9 | 40.7920 | 2.4556 |
8 | 0.08 | 0.92 | 3.0000 | 1.0000 | 1.0000 | 1.2518 | 2964.5 | 41.0361 | 2.4650 |
9 | 0.09 | 0.91 | 3.0000 | 1.0000 | 1.0000 | 1.3007 | 3016.4 | 41.2977 | 2.4741 |
10 | 0.10 | 0.90 | 3.0000 | 1.0000 | 1.0000 | 1.3483 | 3066.9 | 41.5755 | 2.4829 |
|
|
|
|
|
|
|
|
|
|
40 | 0.40 | 0.60 | 3.0000 | 2.0071 | 1.0123 | 1.4285 | 4692.9 | 59.2051 | 3.0899 |
41 | 0.41 | 0.59 | 3.0000 | 2.0575 | 1.0298 | 1.3985 | 4753.4 | 60.1006 | 3.1171 |
42 | 0.42 | 0.58 | 3.0000 | 2.1092 | 1.0480 | 1.3664 | 4814.7 | 61.0222 | 3.1447 |
43 | 0.43 | 0.57 | 3.0000 | 2.1621 | 1.0668 | 1.3323 | 4876.9 | 61.9718 | 3.1728 |
44 | 0.44 | 0.56 | 3.0000 | 2.2164 | 1.0862 | 1.2958 | 4940.1 | 62.9514 | 3.2014 |
45 | 0.45 | 0.55 | 3.0000 | 2.2720 | 1.1064 | 1.2570 | 5004.4 | 63.9635 | 3.2305 |
46 | 0.46 | 0.54 | 3.0000 | 2.3291 | 1.1274 | 1.2158 | 5070.0 | 65.0103 | 3.2601 |
47 | 0.47 | 0.53 | 3.0000 | 2.3877 | 1.1492 | 1.1719 | 5137.0 | 66.0947 | 3.2903 |
48 | 0.48 | 0.52 | 3.0000 | 2.4478 | 1.1718 | 1.1252 | 5205.5 | 67.2194 | 3.3210 |
49 | 0.49 | 0.51 | 3.0000 | 2.5095 | 1.1952 | 1.0758 | 5275.7 | 68.3873 | 3.3522 |
50 | 0.50 | 0.50 | 3.0000 | 2.5729 | 1.2195 | 1.0235 | 5347.7 | 69.6017 | 3.3841 |
51 | 0.51 | 0.49 | 3.0000 | 2.6194 | 1.2400 | 1.0000 | 5414.4 | 70.7405 | 3.4107 |
52 | 0.52 | 0.48 | 3.0000 | 2.6526 | 1.2574 | 1.0000 | 5476.5 | 71.8186 | 3.4333 |
53 | 0.53 | 0.47 | 3.0000 | 2.6862 | 1.2751 | 1.0000 | 5539.2 | 72.9277 | 3.4562 |
54 | 0.54 | 0.46 | 3.0000 | 2.7202 | 1.2932 | 1.0000 | 5602.7 | 74.0695 | 3.4794 |
55 | 0.55 | 0.45 | 3.0000 | 2.7547 | 1.3117 | 1.0000 | 5667.0 | 75.2463 | 3.5030 |
|
|
|
|
|
|
|
|
|
|
85 | 0.85 | 0.15 | 3.0000 | 3.0000 | 3.0000 | 1.0000 | 7395.8 | 121.5131 | 3.9545 |
86 | 0.86 | 0.14 | 3.0000 | 3.0000 | 3.0000 | 1.0005 | 7396.0 | 121.5209 | 3.9546 |
87 | 0.87 | 0.13 | 3.0000 | 3.0000 | 3.0000 | 1.1523 | 7442.5 | 124.1540 | 3.9827 |
88 | 0.88 | 0.12 | 3.0000 | 3.0000 | 3.0000 | 1.3319 | 7498.7 | 127.5789 | 4.0160 |
89 | 0.89 | 0.11 | 3.0000 | 3.0000 | 3.0000 | 1.5455 | 7567.1 | 132.0896 | 4.0556 |
90 | 0.90 | 0.10 | 3.0000 | 3.0000 | 3.0000 | 1.8029 | 7651.9 | 138.1593 | 4.1034 |
91 | 0.91 | 0.09 | 3.0000 | 3.0000 | 3.0000 | 2.1180 | 7759.1 | 146.5312 | 4.1618 |
92 | 0.92 | 0.08 | 3.0000 | 3.0000 | 3.0000 | 2.5106 | 7898.0 | 158.4110 | 4.2346 |
93 | 0.93 | 0.07 | 3.0000 | 3.0000 | 3.0000 | 2.9999 | 8079.3 | 175.4670 | 4.3253 |
94 | 0.94 | 0.06 | 3.0000 | 3.0000 | 3.0000 | 3.0000 | 8079.4 | 175.4708 | 4.3253 |
|
|
|
|
|
|
|
|
|
|
According to the results in Table
Computer simulation is employed to examine the design. The obtained result of energy absorption is
The structure comparison by crash simulation.
In order to examine the efficiency of the generate-first-choose-later method with the hybrid algorithm proposed in this paper, the engineering example is executed 20 times by the computer whose CPU is P8400. All the results in Table
Overall, the method proposed in this paper for multiobjective engineering optimization problems can offer many effective suggestions to designer as a comprehensive reference, and the short computing time speeds up the design.
The proposed generate-first-choose-later method is an effective and efficient approach for multiobjective engineering optimization problems. In the example of crash box, this method gives some valuable reference to design the structure. Relying on the generated reference, the designers can understand the relationship between the design variables and the properties of structures. In addition, the preliminary shape design can be chosen from the optimal solutions.
According to the numerical examples and the engineering example, Algorithm
For the weighted sum approach being adopted, this method has difficulty in searching the solutions when the Pareto curve is not convex. Improving Algorithm
The authors declare that there is no conflict of interests regarding the publication of this paper.
The research is supported by the Fund for the National Natural Science Foundation of China (no. 50975121), the Doctoral Program of Higher Education (no. 20130061120035), the Plan for Scientific and Technology Development of Jilin Province (20130522150JH), and the Fund for Postdoctoral Scientific Research of Jilin Province (RB201337).