Abstract: As artificial agents and robots become more advanced through modern machine learning and complex symbolic methods, testing these systems becomes increasingly difficult. Trained agents must be viewed as a black-box for evaluation, however manual testing of complex agents does not scale well with agent complexity. On the other hand, agents deployed in the real world will encounter novel scenarios not evaluated in experimental settings. To discover edge-case scenarios we propose automatically generating scenarios, where a scenario constitutes both an environment and other simulated agents. We frame searching the space of possible scenarios as a quality diversity (QD) problem, a class of optimization problems that aim to find a collection of solutions spanning the space of specified measure functions, where each solution also maximizes an objective. We then present methods advancing the state-of-the-art of QD algorithms, but also a new problem setting called differentiable quality diversity (DQD) that allows for the objective and measure functions to be first-order differentiable. To address the realism problem, we present methods for representing scenarios via generative models that guarantee task feasibility via mixed integer linear programming. To address expensive evaluation of agents, we demonstrate how deep surrogate models can accelerate environment generation by predicting agent performance and behavior in each proposed environment. Each of these methods is combined into an efficient scenario generation framework that tests and evaluates autonomous agents in a way that can be easily adapted to a diverse array of applications. Finally, I will discuss how advancements in automatic scenario generation research will lead to more robust autonomous systems overall.

Bio: Matt Fontaine is a PhD student at the University of Southern California (USC) advised by Stefanos Nikolaidis. Prior to coming to USC, Matt completed B.S. and M.S. degree at the University of Central Florida (UCF). He also worked as a research assistant in game-based training at UCF’s Institute for Simulation and Training and as a simulation engineer at the autonomous vehicles start-up Drive.ai. His research focuses on the problem of automatic scenario generation for robust human-robot interaction and the algorithmic foundations of quality diversity optimization. Matt’s work has been published in RSS, GECCO, and AAAI, and his work on differentiable quality diversity was recognized with an oral presentation at NeurIPS 2021.