Search Based Software Engineering: Techniques, Taxonomy, Tutorial - Empirical Software Engineering and Verification - page 49

since may algorithms allow for 'human-in-the-loop fitness computation. Such
subjective, human-guided, fitness has been used in interactive evolutionary al-
gorithms for SBSE applied to design-oriented problems [89] and Requirements
Engineering [92].
Use Multiple Fitness Functions: It is not necessary to use only the 'best'
fitness function to guide the search. If the best fitness function is still only the
best at capturing part of the search space there is nothing to prevent the use of
other fitness functions that capture different, perhaps smaller, parts of the solu-
tion space. Fitness functions can be combined in a number of ways to provide an
'agglomerated' fitness function. Both weighting and Pareto optimal approaches
have been widely studied in the SBSE literature [36]. However, using several
fitness functions, each of which applies to different parts of the solutions space
has not been explored in the literature. Given the complicated and heterogenous
nature of many Software Engineering problems, this approach is under-explored
and should receive more attention. In future work on SBSE, we may seek to
bundle patchworks of different fitness functions to solve a single problem, de-
ploying different fitness functions at different times, for different stake holders,
or for different parts of the solutions space.

11.3 My Search Space Is Too Dicult for Search to Work Well

The performance of a search based optimisation algorithms depends crucially
on the search landscape that a fitness function creates when used to guide the
search for a specific problem. If a particular search algorithm performs poorly
on a search space then there are two obvious solutions that immediately present
themselves; do something to modify the way fitness is computed or choose a
different algorithm (one that is better suited to the landscape). In order to take
either course of action, it is important to undertake research into the properties
of the search landscape in order to understand which is the best algorithm to
apply. There has been much work on analysis and characterisation of SBSE
landscapes and fitness functions, but more is required in order to provide a more
complete understanding of the properties to which SBSE is applied.

Analyse Different Fitness Functions: Different characterisations of fitness
can achieve very different results. This has been demonstrated empirically, in
SBSE problems, where the choice of fitness can have different robustness to
noise in the data [51]. The initial choice of fitness function may lead to a search
landscape contains too many plateaux, or other features that make search hard
(needle in a haystack, multimodal features, deceptive basins of attraction etc.).
In these situations, it makes sense to analyse the effect of different fitness func-
tions; each will characterise the problems differently and may have very different
behaviours, even if all agree on the local or global optima.

Use Multiple Fitness Functions: Even if your search problem is inherently
single objective in nature, it may make sense to consider experimenting with
multi objective approaches. It has been shown in previous work on SBSE for