TuneScape: Visualizing Multi-Tuner Parameter Optimization Logs
Authors
Donghee Hong (Sungkyunkwan University), Minjong Kim (Sungkyunkwan University), Sooyoung Cha (Sungkyunkwan University), Jaemin Jo (Sungkyunkwan University)
Presentation
- Session
- Tools of the Trade
- Time
- Thursday, Nov 12, 13:18 – 13:27 (US/Eastern) · session 13:00 – 14:30
- Location
- Hall Essex center
Keywords
Visual analytics, parameter tuning, symbolic execution, hyperparameter optimization, comparative visualization.
Abstract
Tuning symbolic execution engines requires configuring tens of parameters, and researchers often run multiple tuners on one program, producing Multi-Tuner Parameter Optimization Logs (MTPOL). However, existing tools reduce each tuner run to a scalar metric, obscuring where it searches and how its coverage complements other tuners’. We present TuneScape, a map visualization that lays out MTPOL over a shared hexagonal partition of the parameter space, allowing researchers who run multiple tuners to contrast them, inspect the parameter values that drive their differences, and find complementary regions. Two use cases, on a KLEE benchmark and a hyperparameter optimization log, illustrate how TuneScape shows tuner-specific search regions and the parameter values that characterize them, and helps locate cross-tuner combinations that extend branch coverage beyond any single tuner in these cases.
For Practitioners
Software engineers who use symbolic execution engines such as KLEE for automated test generation, along with machine learning practitioners running hyperparameter optimization, are the primary audience for this paper. Practitioners who tune these systems typically run several tuners and compare only their final coverage or accuracy scores, without visibility into where each tuner searched or why one outperformed another. TuneScape lets practitioners inspect each tuner's search region directly, identify the parameter values that explain differences in outcomes, and combine tuners into working sets that cover more branches or samples than any single tuner reaches alone. The tool applies beyond symbolic execution to any setting where multiple tuners search a shared parameter space and report a per-instance outcome, as shown by the hyperparameter optimization use case on XGBoost.