Research map / Reinforcement learning / Policy gradient and actor-critic
RCCDA: Adaptive Model Updates in the Presence of Concept Drift under a Constrained Resource Budget
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RCCDA, a Lyapunov drift-plus-penalty threshold policy that decides when to update a deployed model under concept drift using only past inference loss and a virtual resource queue, with provable budget adherence.
Problem
Adapting deployed ML models to concept drift while strictly meeting a time-averaged resource (update-cost) budget, without expensive drift detection.
Approach
A causal threshold update rule derived from a Lyapunov drift-plus-penalty objective, where the non-causal performance gain is replaced by an estimator built from past inference losses (proportional and derivative terms), and a virtual queue tracks resource debt; provides a convergence bound under KL-bounded drift and a stability bound on constraint violation.
Key results
- On PACS and DigitsDG under an update-rate budget of 0.1, RCCDA achieves the highest average accuracy in all 8 drift-schedule settings (e.g., 72.8% on PACS Burst vs 64.0% for Uniform Random).
- RCCDA's update rate converges toward the 0.1 budget over time while keeping accuracy high.
- Constraint-violation bound goes to zero as T grows under sublinear cumulative drift and bounded update costs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RCCDA instead of Budget-Threshold | PACS · PACS — Burst | accuracy: 67.4 ± 14.9 -> 72.8 ± 4.5 | Helped |
| Use RCCDA instead of Budget-Threshold | PACS · PACS — Step | accuracy: 71.7 ± 8.3 -> 72.0 ± 8.1 | Helped |
| Use RCCDA instead of Uniform | PACS · PACS — Wave | accuracy: 61.8 ± 7.3 -> 67.0 ± 4.7 | Helped |
| Use RCCDA instead of Uniform | PACS · PACS — Spikes | accuracy: 60.8 ± 7.7 -> 73.0 ± 3.6 | Helped |
| Use RCCDA instead of Budget-Threshold | DigitsDG · DigitsDG — Burst | accuracy: 71.9 ± 3.7 -> 77.6 ± 6.3 | Helped |
| Use RCCDA instead of Budget-Increase | DigitsDG · DigitsDG — Step | accuracy: 73.3 ± 5.6 -> 74.6 ± 7.0 | Helped |
| Use RCCDA instead of Uniform | DigitsDG · DigitsDG — Wave | accuracy: 63.6 ± 7.5 -> 65.0 ± 11.2 | Helped |
| Use RCCDA instead of Periodic | DigitsDG · DigitsDG — Spikes | accuracy: 68.1 ± 6.8 -> 74.3 ± 7.3 | Helped |
Limitations
- The practical gain estimator introduces estimation error that can affect policy performance and is not theoretically characterized.
- Theoretical guarantees assume a perfect estimator; the practical policy is an approximation.
- Results reported on a fixed 250-step simulated drift setting with fairly small accuracy gaps on Wave and Step schedules.
Experimental setup
Datasets: PACS DigitsDG OfficeHome MEMD-ABSA
Benchmarks: PACS DigitsDG OfficeHome MEMD-ABSA
Metrics: validation accuracy update rate
Baselines: Uniform Random Periodic Budget-Increase Budget-Threshold
Builds on: Lyapunov drift-plus-penalty framework ADWIN
Closest papers by meaning
- DriftSurf: Stable-State / Reactive-State Learning under Concept Drift (ICML 2021 · similarity 0.88)
- Quilt: Robust Data Segment Selection against Concept Drifts (AAAI 2024 · similarity 0.86)
- Identifying Predictions That Influence the Future: Detecting Performative Concept Drift in Data Streams (AAAI 2025 · similarity 0.85)
- Bayesian Weight Enhancement with Steady-State Adaptation for Test-time Adaptation in Dynamic Environments (ICML 2025 · similarity 0.85)
- Persistent Test-time Adaptation in Recurring Testing Scenarios (NeurIPS 2024 · similarity 0.84)
- When to retrain a machine learning model (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.