ARES: An Autonomous End-to-End Web Test Generation Framework with Proactive Self-Healing

The reliability of automated web test suites is critically un dermined by test script decay—the fragility of manually authored Sele nium or Playwright scripts when user interface (UI) elements change. Existing reactive self-healing tools mitigate individual locator failures but require human-written scripts as a starting point and cannot au tonomously generate complete end-to-end test suites. We introduce ARES, an Autonomous End-to-End Web Test Generation Framework that integrates three machine-learning paradigms: (1) supervised learn ing via YOLOv8 for visual UI element perception, classifying 11 element types with [email protected]=0.87; (2) a Proximal Policy Optimization (PPO) agent for goal-directed web exploration, trained for 10,000 steps across held-out training environments separate from the evaluation SUTs; and (3) a proactive multi-modal self-healing engine using 384-dimensional sentence-transformer embeddings that replaces unstable locators be fore runtime failures occur. ARES also introduces a formal oracle infer ence mechanism that automatically derives three-part test oracles from goal semantics, DOM state transitions, and HTTP response patterns, with precision 94% and recall 100% across injected defects. We evaluate ARES against two baselines—manually authored Se lenium scripts (MS) and a random web crawler (RC)—on three open source web platforms: OpenCart 3.0 (SUT-A), Broadleaf Commerce v6.2.0 (SUT-B), and an Angular enterprise system (SUT-C). In our experiments, ARES reduces test creation time by up to 9× (average 7.9× across SUTs), detects 50% more defects than manual scripting (18 vs. 12 across all SUTs and defect types), reduces runtime self healing overhead by over 90%, improves statement coverage by 12–15 percentage points, and reduces test flakiness from 15–25% to 2–5%. Ablation experiments confirm that each component—visual perception, RL-guided exploration, and proactive self-healing—contributes indepen dently to overall performance. Wilcoxon signed-rank tests confirm sig nificance (p < 0.01) for all metrics. The complete replication package is available at https://doi.org/10.5281/zenodo.18362561.