Reinfier and Reintrainer: Verification and Interpretation-Driven Safe Deep Reinforcement Learning Frameworks
This addresses safety and trustworthiness issues for DRL deployment in real-world applications, representing a novel framework rather than an incremental improvement.
The authors tackled the problem of ensuring verifiable and interpretable safety in deep reinforcement learning (DRL) by proposing Reintrainer, a verification-driven interpretation-in-the-loop framework, and Reinfier, a tool for DRL verification and interpretation, which outperformed state-of-the-art methods on six benchmarks in performance and property guarantees.
Ensuring verifiable and interpretable safety of deep reinforcement learning (DRL) is crucial for its deployment in real-world applications. Existing approaches like verification-in-the-loop training, however, face challenges such as difficulty in deployment, inefficient training, lack of interpretability, and suboptimal performance in property satisfaction and reward performance. In this work, we propose a novel verification-driven interpretation-in-the-loop framework Reintrainer to develop trustworthy DRL models, which are guaranteed to meet the expected constraint properties. Specifically, in each iteration, this framework measures the gap between the on-training model and predefined properties using formal verification, interprets the contribution of each input feature to the model's output, and then generates the training strategy derived from the on-the-fly measure results, until all predefined properties are proven. Additionally, the low reusability of existing verifiers and interpreters motivates us to develop Reinfier, a general and fundamental tool within Reintrainer for DRL verification and interpretation. Reinfier features breakpoints searching and verification-driven interpretation, associated with a concise constraint-encoding language DRLP. Evaluations demonstrate that Reintrainer outperforms the state-of-the-art on six public benchmarks in both performance and property guarantees. Our framework can be accessed at https://github.com/Kurayuri/Reinfier.