I want to make a program which can fiddle with the settings in a VST instrument in order to replicate a particular synth preset on that instrument, given an audio file of the output of the preset.

The end game would be that you could combine a bunch of synthesis vsts, effects, and then large databases of samples (drum loops etc) to regenerate music, allowing very elaborate editing of finished pre-recorded music.

what type of algorithm would you use for this? The training data would be synthesizer configurations and output files.