Safe Autonomy with Algorithmic Improvisation

Speaker Name: 
Daniel Fremont
Speaker Title: 
Ph.D Candidate
Speaker Organization: 
UC Berkeley
Start Time: 
Wednesday, February 13, 2019 - 11:00am
End Time: 
Wednesday, February 13, 2019 - 12:15pm
Location: 
E2-599
Organizer: 
Peter Alvaro

Abstract

Despite massive interest in self-driving cars, the problem of how to ensure the reliability and safety of intelligent autonomous systems remains unsolved. In this talk, I will discuss approaches to safe autonomy based on Algorithmic Improvisation, a framework for automatically synthesizing systems with random but controllable behavior. Algorithmic improvisation can be used in a wide variety of applications where randomness can provide variety, robustness, or unpredictability but safety guarantees or other properties must be ensured. These include software fuzz testing, robotic surveillance, machine music improvisation, human modeling, and generation of synthetic data sets to train and test machine learning algorithms. In this talk, I will discuss both the theory of algorithmic improvisation and its practical applications. I will define the underlying formal language-theoretic problem, “control improvisation”, analyze its complexity and give efficient algorithms to solve it. I will describe in detail two applications to autonomous systems: planning randomized patrol routes for surveillance robots, and generating random scenes of traffic to analyze and improve the reliability of neural networks used for autonomous driving. The latter application involves the design of a domain-specific probabilistic programming language to specify traffic and other scenarios. I will close with prospects for a rigorous design process, driven by algorithmic improvisation and other types of automated formal analysis, to ensure the safety of intelligent autonomous systems.

 

Bio

Daniel Fremont is a PhD student (graduating May 2019) in the Group in Logic and the Methodology of Science at UC Berkeley, working with Sanjit Seshia. He received a B.S. degree in Mathematics and Physics from MIT in 2013. His research is generally in the area of formal methods, using automated reasoning to improve the reliability of software, hardware, and cyber-physical systems. He develops practical algorithms for system design, verification, and testing, as well as theory for the core computational problems underlying them.