Reinforcement learning for model selection applied to a nonlinear dynamical system

Resumo

In the context of digital twins, and the integration of physics-based models with machine learning tools,

this paper proposes a new methodology for model selection and parameter identification, applied to nonlinear dy-
namic problems. Reinforcement learning is used for model selection through Thompson sampling, and parameter

identification is performed using approximate Bayesian computation (ABC). These two methods are applied to-
gether in a one degree-of-freedom nonlinear dynamic model. Experimental data are used in the analysis, and two

different nonlinear models are tested. The initial Beta distribution of each model is updated according to how
successful the model is at representing the reference data (reinforcement learning strategy). At the same time,
the prior Uniform distribution of the model parameters is also updated using a likelihood free strategy (ABC). In
the end, the models’ rewards and the posterior distribution of the parameters of each model are obtained. Several
analyses are made and the potential of the proposed methodology is discussed.