<![CDATA[API-based artificial intelligence (AI) services may experience semantic defects and runtime failures under high-load conditions, necessitating test case prioritization strategies that support the early detection of operational defects. This study aims to evaluate the effectiveness of adaptive test case prioritization using Advantage Actor-Critic (A2C) for AI services under high-load conditions. The experiments were conducted using intent classification services based on DistilBERT and ELECTRA-small, with a test suite derived from Banking77. A2C was compared with random prioritization and performance-based prioritization using full test suite prioritization over ten cycles and five runs, yielding 300 experimental units. Effectiveness was evaluated using Average Percentage of Faults Detected (APFD), Test to First Failure (TTF), and response time. For DistilBERT, A2C achieved the highest mean APFD of 0.883816 and the lowest TTF of 1.000. For ELECTRA-small, A2C achieved the highest mean APFD of 0.674268, with a TTF of 1.000, identical to that of performance-based prioritization. Performance-based prioritization yielded the lowest response times for both models: 3,760.845 ms for DistilBERT and 5,984.625 ms for ELECTRA-small. The Friedman tests indicated statistically significant differences across the three metrics for both models ([test statistics and p-values to be inserted]). Wilcoxon signed-rank tests with Holm correction showed that most method pairs differed significantly. Exceptions included TTF for A2C versus performance-based prioritization on ELECTRA-small, which produced identical results, and response time for A2C versus random prioritization, which did not differ significantly. These findings indicate that A2C's primary advantage lies in its ability to generate effective test case prioritization sequences for earlier operational defect detection rather than achieving the lowest response time.]]>
