Heuristic Analysis from Source Code via Symbolic-Guided Optimization
summary
MetaEase analyzes heuristic source code directly, using symbolic execution (KLEE) to generate seeds per code path and a duality-based gradient for the optimal plus a local Gaussian process surrogate to estimate heuristic gradients, then performs projected gradient ascent within paths. It matches or exceeds MetaOpt across five problem domains and eight heuristics, wins against black-box baselines in 88% of settings, finds previously unknown gaps (examples include >30% demand-pinning loss) and successfully analyzes randomized/nonconvex heuristics like Arrow.
AI-written summary · about
chapters (9)
- 0:00Heuristic analysis intro
- 1:21Heuristic underperformance
- 3:06Limitations of heuristic tools
- 4:33Heuristic analysis challenges
- 5:59MetaEase design and gradient ascent
- 8:49Avoiding local maxima
- 11:57MetaEase workflow
- 12:38Experimental results
- 14:05Conclusion and future
related talks
17mFinding Adversarial Inputs for Heuristics using Multi-level Optimization
47mExploring a complex codebase with AI
26mTowards Efficient Model Comparison Using Automated Program Rewriting
15mHieraSynth: A Parallel Framework for Complete Super-Optimization with Hierarchical(…)
13mMetamorph: Synthesizing Large Objects from Dafny Specifications
30mAn Extensible Feature-Oriented Approach for Fine-Grained Code Quality Analysis
45mWhere AI Meets Code
30mTactic Script Optimisation for Aesop