UC Santa Cruz

We propose a novel training approach for neural networks based on switching among an array of feedback controllers (FC). Traditionally, the neural network training implemented in reinforcement learning problems can be achieved with observable kinematic variables (KV). In this work, our training approach takes a step further by using the metrics-only (MO) or LTM inputs into the networks where each metric and FC fulfills a design specification. Alongside, the reinforcement learning algorithm is a hierarchical control architecture for switching among multiple FCs and its modes. With the designed FCs based on Lyapunov functions or stochastic optimal control, we show that the MO training has a faster convergence with less variations than the one that is based on KV. These results are important for applications requiring large number of controllers. We provide results for a pendulum control problem, a bicycle, and a tricycle navigation problem. In the case of the tricycle problem, we also show that the trained neural network can be applied beyond numerically simulated control problems. The results of this work are illustrated by numerical and virtual reality simulations.