Dissimilar test cases have been proven to be effective to reveal faults in software systems. In the Software Product Line (SPL) context, this criterion has been applied successfully to mimic combinatorial interaction testing in an efficient and scalable manner by selecting and prioritising most dissimilar configurations of feature models using evolutionary algorithms. In this paper, we extend dissimilarity to behavioural SPL models (FTS) in a search-based approach, and evaluate its effectiveness in terms of product and fault coverage. We investigate different distances as well as as single-objective algorithms, (dissimilarity on actions, random, all-actions). Our results on four case studies show the relevance of dissimilarity-based test generation for behavioural SPL models, especially on the largest case-study where no other approach can match it.