A new neural network feature importance method: Application to mobile robots controllers gain tuning

Abstract

This paper proposes a new approach for feature importance of neural networks and subsequently a methodology to determine useful sensor information in high performance controllers, using a trained neural network that predicts the quasi-optimal gain in real time. The neural network is trained using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) algorithm, in order to lower a given objective function. The important sensor information for robotic control are determined using the described methodology. Then a proposed improvement to the tested control law is given, and compared with the neural network’s gain prediction method for real time gain tuning. As a results, crucial information about the importance of a given sensory information for robotic control is determined, and shown to improve the performance of existing controllers.