Stability-Oriented Agentic Reinforcement Learning for Web Decision Systems under Cost and Failure Constraints
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Long-horizon web tasks often exhibit unstable behavior, where small decision errors propagate and lead to budget overruns or catastrophic failures. This study introduces a stability-oriented formulation of agentic reinforcement learning for web automation. Decision stability is quantified by the sensitivity of future cost and failure outcomes to small perturbations in current actions. The learning objective encourages policies that operate in low-sensitivity regions of the decision space while optimizing task success under multiple cost budgets. The proposed approach is evaluated on 1,000–2,000 web tasks with injected UI and latency perturbations. Results show that stability-oriented policies produce smoother trajectories, lower variance in cost consumption, and improved robustness to execution noise compared with standard constrained RL baselines.
Fields of Study & Subjects
Web Data Mining and Analysis
Reinforcement Learning in Robotics
Robotic Process Automation Applications
Publication details
- Journal / source
- American Journal Of Big Data
- Volume
- 7
- Issue
- 1
- Pages
- 202–211
- Publication date
- 25 March 2026
- Language
- EN