Modeling a simple inverted pendulum using a model-based dynamic recurrent neural network

A model-based dynamic recurrent neural network (MBDRNN) is used in this paper to improve the linearized model of a simple inverted pendulum (SIP). The MBDRNN's equations start as those of the linearized SIP model. Then, through back-propagation-based training, the MBDRNN's activation functions' weights are modified with the objective of improving the linearized SIP model. Simulation results show that the MBDRRN effectively improved the linearized model. By tuning several of the MBDRNN parameters, an improved configuration was found yielding a satisfactory' small modeling approximation error.