Towards a Scalable Approach for Risk-Averse Safety Analysis
Undergrad Thesis 2022
University of Toronto
Risk-averse optimal control is an emerging area of research, where the existing risk-averse control algorithms suffer from the curse of dimensionality limiting their practical utility. We propose an approximate yet tractable method, the Monte Carlo Tree Search, to improve the computational efficiency of a current risk-averse exact method with theoretical guarantees. The computational benefits of our proposed algorithm and the cost improvement property it carries are thoroughly explained. We take steps towards showing that the cost improvement property of stochastic rollout still holds in risk-averse settings and obtain a promising result. We discuss the limitations of our algorithm and the potential directions for future implementations.