3.2 Multi-objective optimization
3.4.1 Framework of surrogate-based optimization
The basic strategy of using surrogate models in optimization is quite intuitive. The first
step concerns the generation of an initial sample of vectors which are evaluated with the
true expansive function and used to train a metamodel. Subsequently, the main optimiza-
tion search is invoked. The generation of the sample vector deeply impact the overall
search effectiveness as well as the metamodel accuracy and its predictive capability inside
the design space.
A plethora of strategies have been proposed not only to validate the surrogate, but also
to enhance its accuracy by adding a feedback loop in which the surrogate optimum design
must be confirmed with calls to the true function and used to update the sampling. In so
doing, it is possible to categorize the surrogate-based optimization into two groups:
I Simple-level framework: the optimization is entirely driven by the surrogate model
II Bi-level framework: the true function is used to evaluate the surrogate optimum
designs
These approaches lead to different problems and limitations and are briefly analyzed in
this section.
Simple-level framework
In this context all the solutions have been assessed in the SM and the achieved fitness is
assumed to be comparable to that assessed by the real function. Since prediction with
the SM is much more efficient than that by the expensive analysis code, the optimization
efficiency can be greatly improved. In addition, SMs also serve as an interface between
the analysis code and the optimizer, which make the establishment of an optimization pro-
cedure much easier. The comparison of the conventional and the surrogate-based simple-
level optimization is sketched in Figure 3.6.
(a) (b)
FIGURE
3.6: Comparison of conventional (a) and surrogate-based (b)
simple-level optimization probelm.
This approach is commonly seen in literature ( [88], [89] and [90] just to mention a
few), because the use of a such a simple approach seems to be the most straightforward
in using SM. However, it should be used carefully since its behaviour is highly dependent
on the accuracy of the SM. In fact, when a SM is not properly selected, or it is constructed
with a reduced-size training sample, or the sample is unevenly distributed, the constructed
model will usually be inaccurate. Therefore, if the optimization calculates the entire set
of solutions exclusively with the SM, the entire approach will have more probabilities of
converging to a false optimum, that, in a MOOP, is a Pareto front not corresponding to the
true Pareto front in the real function.
Bi-level framework
It is evident the need to call the true functions (the expensive analysis codes) not only to
validate the optimum solutions of the SM, but also to enhance the accuracy of the SM it-
self, by adding new sample points to the current sampled data set. To this scope, the entire
process can be regarded as a bi-level optimization problem, as sketched in Figure 3.7; the
main optimization concerns the creation and refining of the SMs and needs calls of the
true functions. The sub-optimization uses the current SMs to determine the new sample
sites by using any optimization algorithm such as gradient-based or EA.
F
IGURE3.7: Flowchart of a bi-level surrogate-based optimization.
The selection of new points at which to call the true function, so-called infill points,
represents the heart of the surrogate-based optimization process. Applying a series of infill
points, based on some infill criteria, is also known as adaptive sampling (or updating),
that is we are sampling the objective function in promising areas based on a constantly
changing surrogate. It is therefore important to distinguish between the initial sampling
step which is employed prior to the main optimization search, and the infill sampling step
3.4. Surrogate-based optimization
51
which is performed during the search. Since the infill sampling vectors are iteratively
generated based on an optimization search, the objective functions affect the sampling.
The success or failure of a surrogate-based optimization rests on the correct choice
of model and infill criteria, or in other terms, in the balance between exploration and
exploitation.
Elements of exploration in the infill criterion need the research of global optimum
location. However, pure design space exploration can essentially be viewed as filling
in the gaps between existing sample points. Pure exploration is of dubious merit in an
optimization context [86] because time spent accurately modelling suboptimal regions is
time wasted when all we require is the global optimum itself. Exploration based infill has
its niche in design space visualization and comprehension where the object is to build an
accurate approximation of the entire design landscape to help the designer visualize and
understand the design environment they are working in or when the final SM is to be used
in a realtime control system.
On the contrary, exploitation-based infill criteria are attractive methods for local op-
timization. However, exploiting the surrogate before the design space has been explored
sufficiently, may lead to the global optimum lying undiscovered.
Thus, great caution must be taken in the choice of the infill criteria and a lot different
methods have been proposed; Jones [91] proposed a classification of the infill criteria into
two breeds:
• One-stage approach: the SM is not fixed when calculating the infill criterion,
rather the infill criterion is used to calculate the SM. Goal Seeking and Conditional
Lower Bound are often used
• Two-stage approach: the SM is fitted to the data and the infill criterion calculated
based upon this model: Searching Surrogate Model (SSM), Expected Improvement
and Statistical Lower Bound are the more common
The two-stage approach is far apart the most used. In particular, SSM is the more attractive
because its simplicity and applicability for all the SMs; in this method, an optimizer such
as EA is invoked to find the optimum, which in turn can be employed to refine the SMs.
Such an approach has been found to be very efficient for local exploitation in the design
space. This infill criterion highlights, once again, the importance that EAs have in MOOP
and their capability to fulfil surrogated based approaches.
In document
Metamodel-based design optimization in industrial turbomachinery
(Page 61-63)