The Decision Support Process is a goal-based, weighted, aggregation-based design ranking technique. It is the final step in the optimization, where the optimization is post-processed.
During an optimization, the DesignXplorer optimizer generates a sample set as follows:
Screening: Sample set corresponds to the number of Screening points plus the Min-Max search results. Search results that duplicate existing points are omitted.
Single-objective optimization (NLPQL, MISQP, ASO). Sample set corresponds to the iteration points.
Multiple-objective optimization (MOGA, AMO) Sample set corresponds to the final population.
The Decision Support Process sorts the sample set using the cost function to extract the best candidates. The cost function takes into account both the Importance level of objectives and constraints and the feasibility of points. (The feasibility of a point depends on how constraints are handled. When the Constraint Handling property is set to Relaxed, all infeasible points are included in the sort. When the property is set to Strict, all infeasible points are removed from the sort.) Once the sample set has been sorted, you can change the Importance level and Constraint Handling properties for one or more constraints or objectives without causing DesignXplorer to create more design points. The Decision Support Process sorts the existing sample set again.
Given input parameters,
output parameters, and their individual targets, the collection of
objectives is combined into a single, weighted objective function,
, which is sampled by means of a direct Monte Carlo method using uniform
distribution. The candidate points are subsequently ranked by ascending magnitudes of the
values of
. The function for
(where all continuous input parameters have usable values of type
"Continuous") is given by the following:
(59) |
where:
| |
| |
The normalized objectives (metrics) are:
(60) |
(61) |
where:
| |
| |
| |
| |
| |
The fuzziness of the combined objective function derives from the weights
, which are simply defined as follows:
(62) |
The labels used are defined in Defining Optimization Objectives and Constraints.
The targets represent the desired values of the parameters, and are defined for the continuous input parameters as follows:
(63) |
and, for the output parameters you have the following desired values:
(64) |
where:
| |
| |
| |
Thus, Equation 62 and Equation 63 constitute the input parameter objectives for the continuous input parameters and Equation 62 and Equation 64 constitute the output parameter objectives and constraints.
The following section considers the case where discrete input parameters and continuous input parameters with manufacturable values are possible. Assume that a continuous input parameter with manufacturable values is defined as follows:
(65) |
with one usable value, the following metric is defined:
(66) |
where, as before:
| |
| |
The target value is given by the following:
(67) |
Thus, the GDO objective equation becomes the following (for parameters with discrete usable values).
(68) |
Therefore, Equation 61, Equation 62, and Equation 66 constitute the input parameter objectives for parameters that might be continuous or possess discrete usable values.
The norms, objectives, and constraints as in equations Equation 66 and Equation 67 are also adopted to define the input goals for input parameters of the type Discrete, which are those parameters whose usable alternatives indicate a whole number of some particular design feature (number of holes in a plate, number of stiffeners, and so on.
Thus, equations Equation 61, Equation 62, and Equation 66 constitute the input parameter goals for discrete parameters.
Therefore, the GDO objective function equation for the most general case (where there are continuous and discrete parameters) can be written as the following:
(69) |
where:
|
|
|
|
|
|
From the normed values it is obvious that the lower the value of , the better the design with respect to the desired values and
importance. Thus, a quasi-random uniform sampling of design points is done by a Hammersley
algorithm and the samples are sorted in ascending order of
. The desired number of designs are then drawn from the top of the sorted
list. A crowding technique is employed to ensure that any two sampled design points are
not very close to each other in the space of the input parameters.
Each parameter range is divided into six zones, or rating scales. The location of a
design candidate value in the range is measured according to the rating scales. For
example, for parameter with a range of 0.9 to 1.1, the rating scale for a design candidate
value of 1.0333 is calculated:
[one star] (as 0 indicating neutral, negative values indicating closer
to the target, up to -3. Positive value indicating farther away from the target, up to
+3)
Following the same procedures, you get rating scale for design candidate value of
0.9333 as [two crosses] (away from target). Therefore, the extreme cases are as
follows:
Design Candidate value of 0.9 (the worst), the rating scale is [three crosses]
Design Candidate value of 1.1 (the best), the rating scale is [three stars]
Design Candidate value of 1.0 (neutral), the rating scale is [dash]
Note: Objective-driven parameter values with inequality constraints receive either three stars (the constraint is met) or three red crosses (the constraint is violated).