Robust Regression for Safe Exploration in Control

Abstract

We study the problem of safe learning and exploration in sequential control problems. The goal is to safely collect data samples from operating in an environment, in order to learn to achieve a challenging control goal (e.g., an agile maneuver close to a boundary). A central challenge in this setting is how to quantify uncertainty in order to choose provably-safe actions that allow us to collect informative data and reduce uncertainty, thereby achieving both improved controller safety and optimality. To address this challenge, we present a deep robust regression model that is trained to directly predict the uncertainty bounds for safe exploration. We derive generalization bounds for learning, and connect them with safety and stability bounds in control. We demonstrate empirically that our robust regression approach can outperform the conventional Gaussian process (GP) based safe exploration in settings where it is difficult to specify a good GP prior.

2nd Annual Conference on Learning for Dynamics and Control

References in corpus (9)

• Spectral Normalization for Generative Adversarial Networks
• Safe Model-based Reinforcement Learning with Stability Guarantees
• Neural Lander: Stable Drone Landing Control using Learned Dynamics
• BayesOpt: A Bayesian Optimization Library for Nonlinear Optimization, Experimental Design and Bandits
• Stagewise Safe Bayesian Optimization with Gaussian Processes
• Safe Exploration in Markov Decision Processes
• Conditioning by adaptive sampling for robust design
• Control Regularization for Reduced Variance Reinforcement Learning
• Robust Covariate Shift Prediction with General Losses and Feature Views

Cited by in corpus (5)

• Probabilistic Safety Constraints for Learned High Relative Degree System Dynamics
• ShieldNN: A Provably Safe NN Filter for Unsafe NN Controllers
• Neural-Swarm: Decentralized Close-Proximity Multirotor Control Using Learned Interactions
• Provably Correct Training of Neural Network Controllers Using Reachability Analysis
• Certainty Equivalent Perception-Based Control