SHERPA - An Efficient and Robust Optimization/Search Algorithm
SHERPA - An Efficient and Robust Optimization/Search Algorithm
SHERPA - An Efficient and Robust Optimization/Search Algorithm
Rev. 05.08
Page | 1
SHERPA – An Efficient and Robust Optimization/Search Algorithm
The variables are allowed to vary as follows: This study investigated the ability of each algorithm
to find the known optimal (lowest mass) solution
3 . 0 ≤ H ≤ 7 .0 using a specified number of evaluations. For a given
0.1 ≤ h1 ≤ 1.0 number of allowable evaluations, each algorithm
was run fifty (50) times from different (random)
2.0 ≤ b1 ≤ 12.0 starting conditions. The best solution found in each
0.1 ≤ b2 ≤ 2.0 run was recorded, and the average of these best
solutions was calculated and then normalized by the
known optimal solution. These results are shown in
Figure 2. The standard deviation of these solutions
h1
P was also calculated, as shown in Figure 3.
H b2
Prior to executing the final runs for this study, an
h1 effort was made to “tune” the parameters in each
L
b1
algorithm in order to increase the effectiveness of
the search for this particular problem. The exception
was SHERPA, which has no tunable parameters. All
Figure 1. (a) Cantilever beam with a tip load. (b)
Cross sectional shape variables in the Ibeam. of its parameters are automatically adjusted in an
adaptive manner, as discussed in Section 3.
Five different algorithms were investigated. Four of
In Figure 2, it can be seen that SHERPA required
the algorithms were selected because they are the
many fewer evaluations than the other algorithms
most widely used methods today, and are available
did to find nearly optimal solutions. Alternatively, it
in most commercial optimization packages. The fifth
may be said that for a given number of allowable
algorithm is SHERPA, which is a proprietary method
evaluations, SHERPA found much better designs
available in the software package HEEDS. The
than the other algorithms did. On this particular
algorithms considered are:
problem, SHERPA is evidently much more efficient
SHERPA – simultaneous hybrid exploration that is than the other algorithms considered, even after
robust, progressive and adaptive. This is a multi‐ tuning their parameters to improve performance.
point, hybrid adaptive algorithm described below in Similar conclusions have been reached for many
Section 3. other problems.
GA – genetic algorithm. This is a multi‐point, The results in Figure 3 indicate that SHERPA is also
evolutionary search method that performs global very robust for this problem, having the least
exploration of the design space while searching for amount of variation in the final solutions found for
an optimal solution. It does not require the all levels of evaluations performed. This increased
calculation of solution gradients. robustness is due in part to the hybrid nature of
SHERPA, as described below.
SA – simulated annealing. This is a single‐point
algorithm that is capable of finding global optima but
3. An Overview of SHERPA
often finds a local optimal solution. It is not
dependent on solution gradients. HEEDS contains a unique search strategy called
SHERPA, which stands for Simultaneous Hybrid
NLSQP – non‐linear sequential quadratic Exploration that is Robust, Progressive and Adaptive.
programming. This is a single‐point, gradient‐ based
approach that typically exhibits good convergence During a single search, SHERPA uses multiple search
toward the nearest local optimal solution. methods simultaneously (not sequentially). This
approach takes advantage of the best attributes of
RSM – response surface method. This method each method, and reduces a method’s participation
searches a surrogate approximation model that is in the search if/when it is determined to be
generated by fitting a chosen function to a set of ineffective.
evaluation data points. Here, a quadratic least
squares surface was used.
Page | 2
SHERPA – An Efficient and Robust Optimization/Search Algorithm
Normalized average best solution
2 SHERPA efficiently for many practical engineering design
GA
problems.
SA
1.8
NLSQP
RSM
Aside from being very efficient, robust and effective,
1.6 the main advantages of this approach include:
1.4 • Users need not spend time and effort trying
to understand their design space before
1.2 choosing a suitable algorithm for an
optimization run. SHERPA will learn about
1
50 75 100 150 250 500
the design space and employ the
Maximum allowable evaluations appropriate algorithms as it proceeds
Figure 2. Average best solution found over 50 runs versus toward finding an optimized solution.
the number of allowable evaluations. All solutions are • Users do not need much, if any, expertise in
normalized by the known optimal solution. optimization algorithms and applications,
because SHERPA makes all of the decisions
about which methods to use and how to
0.7 tune them.
Standard deviation of best solution
SHERPA • Users can define a problem realistically,
0.6
GA based on actual engineering or business
0.5 SA costs and benefits, without feeling
NLSQP
RSM
constrained by the capabilities of a
0.4
particular search method. Problem
0.3 definitions can be much broader and
0.2
include a larger number of variables.
0.1
4. Summary
0
50 75 100 150 250 500 The importance of efficiency and robustness in
Maximum allowable evaluations
optimization algorithms was discussed, and the
Figure 3. Standard deviation of the best solutions results of a simple benchmark example were
found over 50 runs versus the number of
allowable evaluations.
described to highlight the performance of some
commonly used algorithms. For this problem, the
A combination of global and local search methods is new hybrid adaptive algorithm SHERPA was shown
used, with the number of different methods used at to be far superior to the other algorithms in terms of
any time ranging between two and ten. both efficiency and robustness.
Each method contains tuning parameters that are
modified automatically during the search according
to knowledge gained about the nature of the design
space. This evolving knowledge about the design
space also determines when and to what extent
each method is used. In other words, SHERPA
efficiently learns about the design space and adapts
itself so as to effectively search all sorts of design
spaces, even very complicated ones. Naturally, there
is no claim that this approach is better for all
problems than some other approach might be, or
that it will always find a global optimal solution, but
it has been shown to work very effectively and
Page | 3