What question did this study set out to answer?

The aim is to develop a framework to evaluate the stability of reinforcement learning policies under stress conditions.

April 6, 2026Open Access

ARCUS-H: Behavioral Stability Under Controlled Stress as a Complementary RL Evaluation Axis

View Full Paper

Ask AI

Bookmark

Authors

Behavioral stability evaluation framework assesses reinforcement learning policies under controlled stress, suggesting new robustness insights.

The aim is to develop a framework to evaluate the stability of reinforcement learning policies under stress conditions.
Developed ARCUS-H framework for post-hoc evaluation of RL policies.
Applied structured perturbations like sensor noise and reward corruption to test agents.
Evaluated agents across five channels of stability: competence, policy consistency, temporal stability, observation reliability, and action entropy.
Generated approximately 1 million evaluation episodes across various environments and algorithms.
Reward accounted for only 5.7% of variance in behavioral stability, indicating limited explanatory power.
SAC agents showed significantly greater fragility compared to TD3 agents when exposed to observation noise.
MuJoCo agents displayed the highest instability while maintaining strong nominal performance.

Karim ZINEBI (2026) studied this question.

View Full Paper

Ask AI

Bookmark

April 6, 2026Open Access

View Full Paper

Ask AI

Bookmark

Authors

Behavioral stability evaluation framework assesses reinforcement learning policies under controlled stress, suggesting new robustness insights.

The aim is to develop a framework to evaluate the stability of reinforcement learning policies under stress conditions.
Developed ARCUS-H framework for post-hoc evaluation of RL policies.
Applied structured perturbations like sensor noise and reward corruption to test agents.
Evaluated agents across five channels of stability: competence, policy consistency, temporal stability, observation reliability, and action entropy.
Generated approximately 1 million evaluation episodes across various environments and algorithms.
Reward accounted for only 5.7% of variance in behavioral stability, indicating limited explanatory power.
SAC agents showed significantly greater fragility compared to TD3 agents when exposed to observation noise.
MuJoCo agents displayed the highest instability while maintaining strong nominal performance.

Karim ZINEBI (2026) studied this question.

View Full Paper

Ask AI

Bookmark