Year: 2024
Type: dissertation
Abstract: The democratization of automated reasoning is a dream for computer scientists like few others.However, despite many years of progress there continue to exist fundamental obstacles in the way.In this thesis we identify one of the key obstacles: the inescapable need for expert help encountered when using automated reasoning tools or algorithms to prove rich sets of properties.This help takes many forms across the many reasoning frameworks that exist, but we argue that in many widely used frameworks the help is a technical intervention that experts are able to come up with by studying a problem, simply using their experience and creativity.In this work we seek to unravel the nature of this creative technical help and take steps towards eliminating it.Although at first sight the workflow we describe may appear informal or unorganized, our study reveals that it is possible to formally characterize the limits of the reasoning power of various automatic tools and heuristics.Furthermore, we show that the expert help in fact bridges (again in a formal sense) the gap between the limits of the reasoning algorithms and the power needed to prove the properties desired.We dub such gaps in automated reasoning creativity gaps and develop new theoretical tools to formally characterize them.Understanding the role of the expert help in a formal sense allows us to formulate welldefined computational problems to solve in order to bridge creativity gaps.We argue that such problems can be solved effectively using a form of learning we call logic learning in this thesis, which refers to the problem of learning logical formulas from rich example structures/logical models.We develop new frameworks and algorithms for logic learning and use them to bridge different creativity gaps.In a third part of our work, we apply the lens of thinking about the dynamics between expert help and automation to interrogate design principles for new verification paradigms that can minimize the kind of user frustrations we focus on in this work, namely dealing with the opaqueness of creativity gaps and the inability to provide the required help to bridge the gaps without deep verification expertise.We formulate the problem of designing a predictable verification framework and develop a new framework for verifying heap manipulating programs that offers a predictable verification experience.We believe that the contributions of this work take significant steps towards eliminating creativity gaps in automated reasoning, and in doing so pave the way further for progress towards the democratization of automated reasoning.1We in fact study this particular problem in this thesis and provide new theoretical results on the limits of popular verification heuristics as well as a formal characterization of the role of expert creative help (see summary of contributions below).2 The contributions in this work have appeared across the following peer-reviewed publications co-authored by the author of this thesis, appearing as citations [15], [16], [17], and [18].
Author: Murali, Adithya
Institution: University of Illinois Urbana-Champaign
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Keywords: automated reasoning, formal verification
Open Access status: green
OpenAlex ID: W7220753027
