Combining Functional and Automata Synthesis to Discover Causal Reactive Programs
summary
Algorithm converts sequences of video frames plus actions into Autumn programs by iterating functional synthesis (fit per-object update functions from an abstract update function matrix) and automata synthesis (hypothesize finite-state latent variables and learn transitions via Sketch or heuristics). It recovers code for 27 of 30 benchmark Autumn programs and 21 of 27 third-party grid-world games, runs in seconds to minutes on traces of hundreds of frames, and exposes limits: requires primitive libraries, handles mostly deterministic dynamics, and sometimes needs more data to succeed.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:23Problem Statement
- 3:34Contributions
- 4:36The Autumn Language
- 11:24Update Function Synthesis
- 12:15Cause Synthesis
- 15:24Automata Synthesis
- 18:41Automata Learning
- 21:00Conclusions
