On-line system identification using context discernment

Mathematical models are often used in system identification applications. The dynamics of most systems, however, change over time and the sources of these changes cannot always be directly determined or measured. To maintain model accuracy, it is desirable to design system identifiers that can adapt to these dynamical shifts. We use reinforcement learning to train an agent to recognize dynamical changes in a modeled system and to estimate new parameter values for the model. The subsequent actions of this agent are characterized as "moving" the parameterized model on an optimal trajectory in model parameter space. It is found that this method is capable of quickly and accurately discerning the correct parameter values.