[legulere]

That sounds like a perfect match for using machine learning. The exact connection what we think is the perfect kerning isn't obvious and the model used by Adobe seems to be arbitrarily selected. Machine learning probably can approximate how a human decides which kerning is best better.

[jack_jennings]

Right, though I'm skeptical that the program would also be able to correctly identify the motivation for the typeface, which drives all of the decisions about how it behaves. That component of the program would end up being far more complex than generating the actual kerning pairs themselves.