xii Preface
functional programming, particulate genes, and neural nets and (most speculatively)
suggests that if the singularity is reachable, it probably will be evolved rather than
autonomously springing into being.
The ninth chapter, titled “Lexicase Selection for Program Synthesis: A Diversity
Analysis,” by Spector and Helmuth, is an exploration of the hypothesis that
lexicase selection improves diversity in a population. Lexicase selection is compared
with tournament selection and implicit fitness sharing. Lexicase showed improved
error diversity, which suggests improved population diversity, thus supporting the
hypothesis and the expected mechanism for lexicase selection.
In the next chapter, “Behavioral Program Synthesis: Insights and Prospects,”
by Krawiec et al., the authors argued at the workshop that a single-valued fitness
function “abuses” program evolution by forcing it to evolve a lump sum of what is
often a complex set of samples. Instead, they propose using an interaction matrix
as a more useful metric as it gives information on specific tests.They argue that
not only is information being “left on the table” with single-valued metrics but that
the overall behavioral characteristic of an evolved solution is lost and a great deal
of nuance and understanding goes missing. They then go on to propose what they
call behavioral synthesis which focuses on the behavior of evolved solutions as the
dominant factor in evolution. This paper suggests that we need a more nuanced
notion of fitness.
The eleventh chapter, “Using Graph Databases to Explore the Dynamics of
Genetic Programming Runs,” McPhee et al. continues the search for understanding
diversity in GP populations, a long-standing focus for research in the GP commu-
nity. However, in this case, the authors are more interested in looking for “critical
moments in the dynamics of a run.” To do this, they use a graph database to
manage the data and then query the database to search for these crucial inflection
points. They focus on the question of whether lexicase selection is truly better than
tournament selection and why this might be. Though a work still in progress, this
chapter suggests that this method of analyzing GP populations is a valuable addition
to the GP toolset and re-raises some of the issues explored in chapter “GP As If You
Meant It: An Exercise for Mindful Practice” by Tozier about looking at the process
and not just the outcome and chapter “Behavioral Program Synthesis: Insights and
Prospects” about the study of behavioral synthesis suggesting that this is an area
where we will see more study in the near future.
The twelfth chapter is titled “Product Choice with Symbolic Regression and
Classification,” by Truscott and Korns. This is one of the first, if not the first use
of GP in market research. Huge amounts of money are spent surveying customers,
and this data is used to predict brand popularity. The authors describe a survey of
cell phones and the analysis produced using the ARC symbolic regression system
adapted to classification. The results show well compared to existing methods and
suggest that more work in this field may be productive.
The thirteenth chapter by Silva et al., is titled “Multiclass Classification Through
Multidimensional Clustering” and revisits the difficult problem of multiclass clas-
sifications using GP. This builds from their earlier work which mapped values into
higher-dimensional space during the training phase and then collected samples into