Search based test generation alone is unable to generate inputs that have complex structures such as dependencies between different data structures.
This paper uses symbolic execution to capture the missing dependencies and then transforms this into a fitness function to guide the search for inputs satisfying the dependencies.
SUSHI implements this for Java and is based on JBSE.
It is evaluated on benchmarks: avl, treemap, caching, tsafe, gantt, clos01, clos72. And compared against JBSE, Seeker and EvoSuite. Results with SUSHI seem to be dramatically better.
Future work is automatic synthesis of invariants.