Absolutely! In the past I worked on an image recognition project where I implemented, tried and thrown away 14 distinct algorithms/approaches (not counting variants!) for a particular problem. TDD is unfit for this because you tend to commit yourself to the set of tests instead of throwing all of them out (all of the approaches were fundamentally different and I had soft real requirements -- so optimized representations were a must). Of course we had some unit tests, but not written in TDD style since the goal was to find the correct approach -- not the correct implementation.
Not 14 different things. 14 different approaches to solve the same image recognition problem. Please do not jump to conclusions. I did not say TDD is not useful at a _later_ stage, but at the research phase it is simply more of a burden than help, because the "refactor" step ends up as "drop everything".
