Clustering-based Approach to Controller Synthesis with Deep Neural Networks

Abstract

This paper addresses control methods for nonlinear dynamical systems. Optimal control is a prominent method for controlling such systems, which involves solving an optimization problem with constraints like system dynamics to determine a series of control inputs. However, nonlinear optimization problems can be time-consuming to solve. Consequently, setting new initial states requires recomputation, leading to inefficiency. To overcome this, methods using previously obtained data to quickly compute control inputs for new initial states have been proposed. Notably, using deep neural networks to learn controllers that output control inputs based on state information has gained attention. However, despite similar control objectives, solutions to the optimization problem can vary significantly with different initial conditions, making learning challenging and reducing estimation accuracy and control performance. This paper proposes a method to improve the estimation accuracy and control performance of controllers by classifying learning data using clustering and then learning based on the classified data. The effectiveness of this approach is verified through numerical experiments on a vehicle obstacle avoidance problem, comparing it with existing methods.