Reproduce “Locally Adaptive Multi-Objective Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jivat Neet Kaur, Isaac Gibbs, Michael I. Jordan

University of California, Berkeley researchers introduce a locally adaptive multi-objective online learning algorithm that modifies the adversarial weight-learning mechanism with an adaptive online method like Fixed Share. This approach consistently achieves low multiaccuracy error over local time intervals while preserving predictive accuracy, outperforming non-adaptive and other adaptive baselines on energy forecasting and recidivism prediction tasks.

Key insights

  • Modifying the weight-learning mechanism with an adaptive online algorithm (e.g., Fixed Share) is empirically more effective for achieving local adaptivity than augmenting the objective set itself.
  • Including a dedicated prediction error objective is crucial for preventing pathological predictions and ensuring that improving multi-objective criteria does not degrade overall predictive accuracy relative to a baseline.
  • Adaptive tuning of the learning rate and the 'share' parameter enables robust and rapid adaptation to arbitrary distribution shifts, providing consistent performance over diverse local time intervals.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.