Modeling Black-Box Components with Probabilistic Synthesis
summary
Presyn builds executable models of opaque library functions by composing a library of 11 fragment families, scoring fragments with a probabilistic fragment classifier and ranking fragment compositions with a Markov model, then filling holes by enumerative refinement. Evaluation uses a 112-function benchmark compiled to run across five synthesizers; the best prior system reaches about 50% when given handmade priors, while Presyn requires minimal prior information and significantly outperforms the others in coverage and practical portability use cases.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:28BLACK BOXES
- 1:23EXAMPLE
- 1:59MINIMAL PROBLEM STATEMENT
- 3:50TESTING
- 4:55COMPOSITIONAL SOLUTIONS
- 6:08PROBABILISTIC MODELS
- 6:36WHICH FRAGMENTS?
- 7:23WHICH COMPOSITION?
- 8:04TRAINING
- 9:06SYNTHESIS
- 10:31BENCHMARKING
- 14:11INSIGHTS
- 15:05USES
related talks
14mSequential Monte Carlo Program Synthesis with Refinement Proposals
38mProgram Synthesis: A Dream Realized?
43mThe Power of Probabilistic Programming
17mProgrammatic Strategy Synthesis: Resolving Nondeterminism in Probabilistic Progr...
38mProbabilistic programming and meta-programming in Clojure
01h 10mModels and Programs: Better Together
01h 03mAn Overview of Probabilistic Programming
47mBig Ideas in Program Synthesis