Abstract:
This paper presents a method for precise prediction of the behavior of an inverted pendulum on a cart system. We have improved the accuracy of prediction beyond what can be achieved through traditional model-based simulation. This improvement has been achieved through learning of the differences between simulation and experimental results. Specifically, a three layered neural network known in the literature as denoising autoencoder has been used for learning. The proposed method consists of three steps. First step is to design linear controller for the inverted pendulum using text book methods and perform simulations. Second step is to perform experiments on the actual hardware of the inverted pendulum in the laboratory using the same controller as in first step. Third step is to learn the difference between simulation results and the results from the experiments using neural networks. Now the learned neural network is used to predict lab experiment results based on simulations with different initial conditions and reference values than the ones used to train the network. We have designed Linear Quadratic Regulator for demonstration of the proposed method. Results from the autoencoder have been reported. It is found that the autoencoder can predict the actual behavior of the pendulum with reasonable accuracy.
