Reinforcement learning and approximate Bayesian computation (RL-ABC) for model selection and parameter calibration of time-varying systems

This paper extends the recently developed methodology for model selection and parameter identification called RL-ABC (Ritto et al., 2022) (reinforced learning and approximate Bayesian computation) to time-varying systems. To tackle slowly-varying systems and detect abrupt changes, new features are proposed. (1) The probability of sampling the worst model has now a lower bound; because it cannot disappear, once it might be useful in the future as the system evolves. (2) A memory term (sliding window) is introduced such that past data can be forgotten whilst updating the reward; which might be useful depending on how fast the system changes. (3) The algorithm detects a change in the system by monitoring the models' acceptance; a significant drop in acceptance indicates a change. If the system changes the algorithm is reset: new parameter ranges are computed and the rewards are restarted. To test the proposed strategy, new experimental data is obtained from a test rig with non-linear restoring force characteristics. The amplitude of the dynamical experiment is obtained with the control -based continuation strategy varying the excitation amplitude, and three Duffing-like models are used to represent the system. The results are consistent, and the strategy is able to detect changes and update parameter estimation and model predictions.