Paper

SDPERL: A Framework for Software Defect Prediction Using Ensemble Feature Extraction and Reinforcement Learning

M. Hesamolhokama*, A. Shafiee*, M. Ahmaditeshnizi*, M. Fazli, J. Habibi

arXiv preprintarXiv:2412.079272024cited 13×co-first author

Abstract

We frame software defect prediction as a sequential decision problem and pair an ensemble of feature extractors with a Proximal Policy Optimization (PPO) agent. The agent learns to weight candidate features and route classifier proposals, lifting accuracy on standard benchmarks while remaining interpretable about which features each class of defect responds to.

Figure 1 — interactive

Hover the chart to compare values.

Hover the curves. Three conditions: baseline, ensemble-only, PPO + ensemble. The agent learns to favor features that respond cleanly to the defect class; accuracy climbs once policy weight crosses ~0.6.

Methods

—Ensemble feature extraction
—PPO
—Reinforcement learning

Where it was done

HABIBI Laboratory, Sharif.

Cite

@misc{sdperl_2024,
  title         = {{SDPERL: A Framework for Software Defect Prediction Using Ensemble Feature Extraction and Reinforcement Learning}},
  author        = {M. Hesamolhokama and A. Shafiee and M. Ahmaditeshnizi and M. Fazli and J. Habibi},
  year          = {2024},
  howpublished  = {arXiv preprint arXiv:2412.07927},
  archivePrefix = {arXiv},
  eprint        = {2412.07927},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2412.07927v1}
}