This study introduces a novel two-stage method, GPAction, for detecting environment drift in reinforcement learning settings. We first train a Gaussian process predicting the reinforcement learning agents' actions and then detect environment drifts by monitoring the mean squared error between the predicted actions of a Gaussian process action predictor and actual actions. Our proposed method is evaluated against three baselines across four environments with continuous action spaces. Results demonstrate the superior performance of GPAction in detecting environment drift. In an ablation study, by analyzing the plots and AUC values of the MSEs, we show that our method GPAction can provide more distinguishable monitoring metrics than the Gaussian process state predictor.

Top