search

mlb2251 / stitch

A scalable abstraction learning library
MIT License
74 stars 8 forks source link

Two ways to tighten upper bound #139

Open mlb2251 opened 2 years ago

mlb2251 commented 2 years ago

Two alternatives that I think might be pretty effective at tightening the bound:

  1. We can tighten the local upper bound at each location using any arguments that have already been decided. I think this information might already be freely available so it might be cheap. Even though final abstractions tend to take very small arguments, bigger abstractions might not
  2. If you dont do the above bound tightening, then the bound is equal to the subtree size meaning ancestor match locations subsume lower ones so you can discard the lower ones and not include them in the bound since they cant improve it any more than "delete the parent tree"
mlb2251 commented 1 year ago

honestly curious if this will make it actually worse as a heuristic bc it correlates less w match locs?