Edward Kim

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2023-214

August 11, 2023

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-214.pdf

Cyber-physical systems (CPS) are increasingly becoming autonomous. Self-driving cars, for example, need to navigate through bustling streets of San Francisco. To avoid the risk of injuries, simulation provides a safe setting to extensively test these systems prior to their deployment. However, as these autonomous systems operate in more complex environments, this poses a challenge as to how to formally model and generate these environments in simulation. In this dissertation, we argue that a domain-specific probabilistic programming language (PPL) can be an adequate formalism to naturally capture the stochasticity and the constraints deriving from physical interactions of these systems with their environments.

The contribution of this thesis is to show how a domain-specific PPL can be effectively used as an environment modeling formalism to test, validate, and train autonomous systems. First, we formalize a machinery to scalably test a system of multi-objectives in parallel simulations with distributions of environments, and summarize the likely causes of those failures in an interpretable manner. Second, we validate whether the identified system failures in simulation transfer to reality, with probabilistic programs as consistent environment models across both simulation and reality. Informed by the validated system failures, we develop algorithms to train components of autonomous systems to be robust against those failures. Furthermore, we devise a personalized algorithm to also train human-CPS (h-CPS), cyber-physical systems that operate in concert with human operators, via simulations in augmented and virtual reality.

Advisors: Alberto L. Sangiovanni-Vincentelli and Sanjit A. Seshia

---

BibTeX citation:

@phdthesis{Kim:EECS-2023-214,
    Author= {Kim, Edward},
    Title= {Simulation-Based Testing, Validation, and Training with Probabilistic Programming},
    School= {EECS Department, University of California, Berkeley},
    Year= {2023},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-214.html},
    Number= {UCB/EECS-2023-214},
    Abstract= {Cyber-physical systems (CPS) are increasingly becoming autonomous. Self-driving cars, for example, need to navigate through bustling streets of San Francisco. To avoid the risk of injuries, simulation provides a safe setting to extensively test these systems prior to their deployment. However, as these autonomous systems operate in more complex environments, this poses a challenge as to how to formally model and generate these environments in simulation. In this dissertation, we argue that a domain-specific probabilistic programming language (PPL) can be an adequate formalism to naturally capture the stochasticity and the constraints deriving from physical interactions of these systems with their environments. 

The contribution of this thesis is to show how a domain-specific PPL can be effectively used as an environment modeling formalism to test, validate, and train autonomous systems. First, we formalize a machinery to scalably test a system of multi-objectives in parallel simulations with distributions of environments, and summarize the likely causes of those failures in an interpretable manner. Second, we validate whether the identified system failures in simulation transfer to reality, with probabilistic programs as consistent environment models across both simulation and reality. Informed by the validated system failures, we develop algorithms to train components of autonomous systems to be robust against those failures. Furthermore, we devise a personalized algorithm to also train human-CPS (h-CPS), cyber-physical systems that operate in concert with human operators, via simulations in augmented and virtual reality.},
}

EndNote citation:

%0 Thesis
%A Kim, Edward 
%T Simulation-Based Testing, Validation, and Training with Probabilistic Programming
%I EECS Department, University of California, Berkeley
%D 2023
%8 August 11
%@ UCB/EECS-2023-214
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-214.html
%F Kim:EECS-2023-214