Test-based problems are search and optimization problems in which candidate solutions interact with multiple tests (examples, fitness cases, environments) in order to be evaluated. The approach conventionally adopted in most search and optimization algorithms involves aggregating the interaction outcomes into a scalar objective. However, passing different tests may require unrelated `skills' that candidate solutions may vary on.Scalar tness is inherently incapable of capturing such di erences and leaves a search algorithm largely uninformed about the diverse qualities of individual candidate solutions. In this paper, we discuss the implications of this fact and present a range of methods that avoid scalarization by turning the outcomes of interactions between programs and tests into 'search drivers' - partial, heuristic, transient pseudo-objectives that form multifaceted characterizations of candidate solutions. We demonstrate the feasibility of this approach by reviewing the experimental evidence from past work, confront it with related research endeavors, and embed it into a broader context of behavioral program synthesis.
Acceleration of Genetic Programming by Hierarchical Structure Learning: A Case Study on Image Recognition Program Synthesis
