I like this idea a lot, and have had a similar idea before. It seems like a good idea to try multiple different testing strategies vs. looking for one optimal one, since it doesn’t look like a general optimal one exists at the moment.
So it’s probably a good idea to just combine swarm testing, unaltered randomized testing, and even approaches with hybrid manual / generated inputs like the category partition method.
This swarm testing idea is also probably a realistic distribution of usage. It’s not like every user uses every single action in an application. Certain user archetypes probably cluster into subsets of features.
Probably. Cool idea.
Also, Directed Swarm Testing.
If you want to focus on a specific part of the program, you can also automatically abstract the input interface/APIs and use the abstraction as a template for further inputs.