# Non-linear Goal Programming Using Multi-Objective Genetic ... programming or minimax goal...

date post

20-Mar-2020Category

## Documents

view

0download

0

Embed Size (px)

### Transcript of Non-linear Goal Programming Using Multi-Objective Genetic ... programming or minimax goal...

Non-linear Goal Programming Using Multi-Objective Genetic Algorithms

Kalyanmoy Deb�

Kanpur Genetic Algorithms laboratory (KanGAL) Department of Mechanical Engineering Indian Institute of Technology, Kanpur

Kanpur, PIN 208 016, India E-mail: deb@iitk.ac.in

Technical Report No. CI-60/98 October 1998

Department of Computer Science/XI University of Dortmund, Germany

Abstract

Goal programming is a technique often used in engineering design activities primarily to find a com- promised solution which will simultaneously satisfy a number of design goals. In solving goal program- ming problems, classical methods reduce the multiple goal-attainment problem into a single objective of minimizing a weighted sum of deviations from goals. Moreover, in tackling non-linear goal programming problems, classical methods use successive linearization techniques, which are sensitive to the chosen starting solution. In this paper, we pose the goal programming problem as a multi-objective optimization problem of minimizing deviations from individual goals. This procedure eliminates the need of having ex- tra constraints needed with classical formulations and also eliminates the need of any user-defined weight factor for each goal. The proposed technique can also solve goal programming problems having non- convex trade-off region, which are difficult to solve using classical methods. The efficacy of the proposed method is demonstrated by solving a number of non-linear test problems and by solving an engineering design problem. The results suggest that the proposed approach is an unique, effective, and most practical tool for solving goal programming problems.

Keywords: Goal programming, Genetic algorithms, Engineering design

1 Introduction

Developed in the year 1955, goal programming method has enjoyed innumerable applications in engineering design activities1�5. Goal programming is different in concept from non-linear programming or optimization techniques in that the goal programming attempts to find one or more solutions which satisfy a number of goals to the extent possible. Instead of finding solutions which absolutely minimize or maximize objective functions, the task is to find solutions that, if possible, satisfy a set of goals, otherwise, violates the goals minimally. This makes the approach more appealing to practical designers compared to optimization methods.

The most common approach to classical goal programming techniques is to construct a non-linear pro- gramming problem (NLP) where a weighted sum of deviations from targets is minimized 6. The NLP problem also contain a constraint for each goal, restricting the corresponding criterion function value to be within the specified target values. A major drawback with this approach is that it requires the user to specify a set of weight factors, signifying the relative importance of each criterion. This makes the approach subjective to the user. Moreover, the weighted goal programming approach has difficulty in finding solutions in problems having non-convex feasible decision space. Although there exists other methods such as lexicographic goal

�Presently visiting Computer Science Department/LS11, University of Dortmund, Germany (deb@ls11.informatik.uni- dortmund.de)

1

programming or minimax goal programming6;7, these methods are also not free from the dependence on the relative weight factor for each criterion function.

In this paper, we suggest using a multi-objective genetic algorithm (GA) to solve the goal programming problem. In order to use a multi-objective GA, each goal is converted into an equivalent objective function. Unlike the weighted goal programming method, the proposed approach does not add any artificial constraint into its formulation. Since, multi-objective GAs have been shown to find multiple Pareto-optimal solutions 8;9, the proposed approach is likely to find multiple solutions to the goal programming problem, each correspond- ing to a different setting of the weight factors. This makes the proposed approach independent from the user. Moreover, since no explicit weight factor for each criterion is used, the method is also not likely to have any difficulty in finding solutions for problems having non-convex feasible decision space.

It is worthwhile to highlight here that the use of a multi-objective optimization technique to solve goal programming problems is not new8;9 and is novel. But the inefficiency of classical non-linear multi-objective optimization methods has led the researchers and practitioners to only concentrate on solving linear goal programming problems. Although there are some attempts to use sequential linear goal programming ap- proaches, where a linear approximation of the non-linear problem is solved sequentially, the methods have not been successful6. Multi-objective GAs are around for last five years or so and have been shown to solve various non-linear multi-objective optimization problems successfully 10�13. As a result of these interests, there exist now a number of multi-objective GA implementations 8;9;14�16. In this paper, we show how one such GA implementation can make the non-linear goal programming easier and practical to use.

In the remainder of the paper, we briefly discuss the concept of goal programming and argue why classical goal programming methods are not adequate tools. Thereafter, we present the working principle of one multi- objective GA implementation—non-dominated sorting GA (NSGA). The usefulness of the proposed approach is demonstrated by solving five different test problems and an engineering design problem using NSGA.

2 Goal Programming

Goal programming was first introduced in an application of a single-objective linear programming problem by Charnes, Cooper, and Ferguson1. However, goal programming gained popularity after the works of Ignizio 4, Lee17, and others. Romero6 presented a comprehensive overview and listed a plethora of engineering ap- plications where goal programming technique has been used. The main idea in goal programming is to find solutions which attain a pre-defined target for one or more criterion function. If there exists no solution which achieves targets in all criterion functions, the task is to find solutions which minimize deviations from tar- gets. Goal programming is different from non-linear programming problems (NLPs), where the main idea is to find solutions which optimizes one or more criteria 18;19. There is no concept of a goal in a mathemati- cal programming problem. We illustrate the concept of goal programming by considering a single-criterion problem.

Let us consider a design criterion f(~x), which is a function of a solution vector ~x. In the context of NLP, the objective is to find the solution vector ~x � which will minimize or maximize f(~x). Without loss of generality, we consider criterion functions which is to be minimized (such as fabrication cost of an engineering component). In most design problems, there exists a number of constraints which make a certain portion (~x 2 F ) of the search space feasible. It is imperative that the optimal solution ~x � is feasible, that is, ~x� 2 F . In a goal programming, a target value t is chosen for every design criterion. One of the design goals may be to find a solution which attains a cost of t:

goal (f(~x) = t) ; ~x 2 F : (1)

If the target cost t is smaller than the minimum possible cost f(~x �), naturally there exists no feasible solution which will attain the above goal exactly. The objective of goal programming is then to find that solution which will minimize the deviation d between the achievement of goal and the aspiration target, t. The solution for this problem is still ~x� and the overestimate is d = f(~x�) � t. Similarly, if target cost t is larger than the maximum feasible cost fmax, the solution of the goal programming problem is ~x which makes f(~x) = fmax.

2

However, if the target cost t is within [f(~x�); fmax], the solution to the goal programming problem is that feasible solution ~x which makes the criterion value exactly equal to t. Although this solution may not be the optimal solution of the constrained f(~x), this solution is the outcome of the above goal program.

In the above example, we have considered a single-criterion problem. Goal programming is hardly used for single criterion problems. In fact, goal programming brings interesting scenarios when multiple criteria are considered. In the above example, an ‘equal-to’ type goal is discussed. However, there can be four different types of goal criteria, as shown below7:

1. Less-than-equal-to (f(~x) � t),

2. Greater-than-equal-to (f(~x) � t),

3. Equal-to (f(~x) = t), and

4. Within a range (f(~x) 2 [tl; tu]).

In order to tackle above goals, usually two non-negative deviation variables (n and p) are introduced. For the less-than-equal-to type goal, the positive deviation p is subtracted from the criterion function, so that f(~x) � p � t. Here, the deviation p quantifies the amount by which the criterion value has surpassed the target t. The objective of goal programming is to minimize the deviation p so as to find the solution for which the deviation is minimum. If f(~x) > t, the deviation p should take a non-zero positive value, otherwise it must be zero. For the greater-than-equal-to type goal, a negative deviationn is added to the criterion function, so that f(~x) + n � t. The deviation n quantifies the amount by which the criterion function has not satisfied the target t. Here, the objective of goal programming is to minimize the deviation n. For f(~x) < t, the deviation n should take a nonzero positive value, otherw

Recommended

*View more*