Intelligent Resilience Testing for Decision-Making Agents with Dual-Mode Surrogate Adaptation

Testing and evaluating decision-making agents remains challenging due to unknown system architectures, limited access to internal states, and the vastness of high-dimensional scenario spaces. Existing testing approaches often rely on surrogate models of decision-making agents to generate large-scale scenario libraries; however, discrepancies between surrogate models and real decision-making agents significantly limit their generalizability and practical applicability. To address this challenge, this paper proposes intelligent resilience testing (IRTest), a unified online adaptive testing framework designed to rapidly adjust to diverse decision-making agents. IRTest initializes with an offline-trained surrogate prediction model and progressively reduces surrogate-to-real gap during testing through two complementary adaptation mechanisms: (i) online neural fine-tuning in data-rich regimes, and (ii) lightweight importance-sampling-based weighting correction in data-limited regimes. A Bayesian optimization strategy, equipped with bias-corrected acquisition functions, guides scenario generation to balance exploration and exploitation in complex testing spaces. Extensive experiments across varying levels of task complexity and system heterogeneity demonstrate that IRTest consistently improves failure-discovery efficiency, testing robustness, and cross-system generalizability. These results highlight the potential of IRTest as a practical solution for scalable, adaptive, and resilient testing of decision-making agents.

💡 Research Summary

This paper addresses the critical challenge of efficiently testing and validating complex “black-box” decision-making agents, such as those used in autonomous driving and robotics. The core difficulty lies in the agents’ unknown internal architectures, limited observability of internal states, and the vastness of high-dimensional scenario spaces. While intelligent testing methods often use surrogate models to predict failure-prone scenarios and generate test libraries, a significant “surrogate-to-real gap” emerges when the surrogate model differs from the actual agent under test, severely limiting generalizability and practical utility.

To bridge this gap, the authors propose Intelligent Resilience Testing (IRTest), a unified online adaptive testing framework. IRTest starts with an offline-trained Surrogate Prediction Model (SPM) – typically a neural network – that estimates the failure probability of a surrogate agent. The key innovation is that IRTest progressively reduces the surrogate-to-real gap during the testing process itself by continuously adapting the SPM using real-time feedback from testing the actual target agent.

The framework features two complementary adaptation mechanisms tailored to different testing resource constraints:

1. Data-Rich Regime Adaptation: When sufficient testing interactions are possible, the neural SPM is online fine-tuned. The weights of the network are updated directly using newly collected failure/success labels, allowing the model to learn and correct its predictions for the specific target agent.
2. Data-Limited Regime Adaptation: When testing interactions are scarce and expensive, IRTest employs a lightweight importance-sampling-based weighting correction. Instead of fine-tuning a single model, it uses a mixture of pre-trained SPMs. The framework efficiently adjusts the combination coefficients of these models based on a small amount of real test data, quickly aligning the mixture’s predictions with the target agent’s behavior without costly model retraining.

The selection of which scenario to test next is guided by a bias-corrected Bayesian Optimization (BO) strategy. A novel acquisition function is designed that incorporates a correction term for the prediction bias of the currently adapted SPM. This enables IRTest to intelligently balance exploration (searching for new, unknown failure regions) and exploitation (concentrating tests on predicted high-risk areas), with the balance shifting as the surrogate model becomes more accurate.

The paper presents extensive experiments across varying levels of task complexity and system heterogeneity (e.g., testing different autonomous driving planning algorithms). The results demonstrate that IRTest consistently and significantly outperforms conventional non-adaptive surrogate-based testing methods and other online adaptation baselines. Key performance improvements include:

• Higher Failure Discovery Efficiency: More failures found per unit of testing effort.
• Enhanced Testing Robustness: Stable performance across different target agents and random seeds.
• Superior Cross-System Generalizability: Effective adaptation from a surrogate agent to a distinctly different target agent.

In conclusion, IRTest provides a scalable, adaptive, and resilient methodology for testing complex decision-making agents. By dynamically closing the surrogate-to-real gap online, it offers a practical pathway toward reliable validation of AI systems in safety-critical applications, even with limited real-world testing resources and facing unknown agent architectures.