Introduction
Most realistic optimization problems, particularly those in design, require the simultaneous optimization of more than one objective function. Some examples:
- In bridge construction, a good design is characterized by low total mass and high stiffness.
- Aircraft design requires simultaneous optimization of fuel efficiency, payload, and weight.
- In chemical plant design, or in design of a groundwater remediation facility, objectives to be considered include total investment and net operating costs.
- A good sunroof design in a car could aim to minimize the noise the driver hears and maximize the ventilation.
- The traditional portfolio optimization problem attempts to simultaneously minimize the risk and maximize the fiscal return.
In these and most other cases, it is unlikely that the different objectives would be optimized by the same alternative parameter choices. Hence, some trade-off between the criteria is needed to ensure a satisfactory design.
Multicriteria optimization has its roots in late-nineteenth-century welfare economics, in the works of Edgeworth and Pareto. A mathematical description is as follows:
where n >= 2 and
denotes the feasible set constrained by equality and inequality constraints and explicit variable bounds. The space in which the objective vector belongs is called the objective space and image of the feasible set under F is called the attained set.
The scalar concept of ``optimality'' does not apply directly in the multiobjective setting. A useful replacement is the notion of Pareto optimality. Essentially, a vector is said to be Pareto optimal for (MOP) if all other vectors have a higher value for at least one of the objective functions , or else have the same value for all objectives. Formally speaking, we have the following definition:
A point is said to be (glob ally) Pareto optimal or a (globally) efficient solution or a non-dominated or a non-inferior point for (MOP) if and only if there is no such that for all , with at least one strict inequality.
Pareto optimal points are also known as efficient, non-dominated, or non-inferior points.
We can also speak of locally Pareto optimal points, for which the definition is the same as the one just given, except that we restrict attention to a feasible neighborhood of . That is, if denotes a ball of radius around the point
We can also speak of locally Pareto optimal points, for which the definition is the same as the one just given, except that we restrict attention to a feasible neighborhood of . That is, if denotes a ball of radius around the point , we require that for some , there is no such that
with at least one strict inequality.
Typically, there is an entire curve or surface of Pareto points, whose shape indicates the nature of the tradeoff between different objectives.
Solution Techniques
The multiobjective problem is almost always solved by combining the multiple objectives into one scalar objective whose solution is a Pareto optimal point for the original MOP. Most algorithms have been developed in the linear framework (i.e. linear objectives and linear constraints), but the techniques described below are also applicable to nonlinear problems.
Minimizing Weighted Sums of Functions
A standard technique for MOP is to minimize a positively weighted convex sum of the objectives, that is,
It is easy to prove that the minimizer of this combined function is Pareto optimal. It is up to the user to choose appropriate weights. Until recently, considerations of computational expense forced users to restrict themselves to performing only one such minimization. Newer, more ambitious approaches aim to minimize convex sums of the objectives for various settings of the convex weights, therefore generating various points in the Pareto set. Though computationally more expensive, this approach gives an idea of the shape of the Pareto surface and provides the user with more information about the trade-off among the various objectives. However, this method suffers from two drawbacks. First, the relationship between the vector of weights and the Pareto curve is such that a uniform spread of weight parameters rarely produces a uniform spread of points on the Pareto set. Often, all the points found are clustered in certain parts of the Pareto set with no point in the interesting ``middle part'' of the set, thereby providing little insight into the shape of the trade-off curve. The second drawback is that non-convex parts of the Pareto set cannot be obtained by minimizing convex combinations of the objectives (note though that non-convex Pareto sets are seldom found in actual applications).
Homotopy Techniques
Homotopy techniques aim to trace the complete Pareto curve in the bi-objective case (n=2). By tracing the full curve, they overcome the sampling deficiencies of the weighted-sum approach. The main drawback is that this approach does not generalize to the case of more than two objectives. For more information, see Rao and Papalambros [7] and Rakowska, Haftka, and Watson [6].
Goal Programming
In the goal programming approach, we minimize one objective while constraining the remaining objectives to be less than given target values. This method is especially useful if the user can afford to solve just one optimization problem. However, it is not always easy to choose appropriate ``goals'' for the constraints. Goal programming cannot be used to generate the Pareto set effectively, particularly if the number of objectives is greater than two.
Normal-Boundary Intersection (NBI)
The normal-boundary intersection method uses a geometrically intuitive parametrization to produce an even spread of points on the Pareto surface, giving an accurate picture of the whole surface. Even for poorly scaled problems (for which the relative scalings on the objectives are vastly different), the spread of Pareto points remains uniform. Given any point generated by NBI, it is usually possible to find a set of weights such that this point minimizes a weighted sum of objectives, as described above. Similarly, it is usually possible to define a goal programming problem for which the NBI point is a solution. NBI can also handle problems where the Pareto surface is discontinuous or non-smooth, unlike homotopy techniques. Unfortunately, a point generated by NBI may not be a Pareto point if the boundary of the attained set in the objective space containing the Pareto points is nonconvex or `folded' (which happens rarely in problems arising from actual applications).
NBI requires the individual minimizers of the individual functions at the outset, which can also be viewed as a drawback.
NBI was developed by Das and Dennis ([8], [1]). A public domain Matlab 4.2 implementation of NBI is available here.
Multilevel Programming
Multilevel programming is a one-shot optimization technique and is intended to find just one ``optimal'' point as opposed to the entire Pareto surface. The first step in multilevel programming involves ordering the objectives in terms of importance. Next, we find the set of points for which the minimum value of the first objective function is attained. We then find the points in this set that minimize the second most important objective. The method proceeds recursively until all objectives have been optimized on successively smaller sets.
Multilevel programming is a useful approach if the hierarchical order among the objectives is of prime importance and the user is not interested in the continuous trade-off among the functions. However, problems lower down in the hierarchy become very tightly constrained and often become numerically infeasible, so that the less important objectives have no influence on the final result. Hence, multilevel programming should surely be avoided by users who desire a sensible compromise solution among the various objectives.