Title:
An Information-Theoretic Framework for Resource-Aware Abstraction and Planning for Autonomous Agents
An Information-Theoretic Framework for Resource-Aware Abstraction and Planning for Autonomous Agents
Author(s)
Larsson, Daniel T.
Advisor(s)
Tsiotras, Panagiotis
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
In the modern era of autonomy, autonomous systems have seen deployment in a number of both terrestrial and extraterrestrial applications including drone delivery systems, warehouse robotics, aerial surveillance, self-driving cars, and mars exploration. However, the systems deployed in the aforementioned applications differ in their size, sensing ability, on-board information-processing resources, as well as their communication capabilities. Consequently, to develop the next-generation of independent and self-sufficient intelligent systems, frameworks that endow autonomous systems the ability to tailor their information processing for decision-making, planning, and perception, in accordance with their on-board resources in a task-specific manner is of paramount importance. For this reason, we consider in this dissertation, the development of approaches for resource-aware, task-driven abstraction in autonomous systems.
The process of abstraction, or equivalently, the identification of relevant and irrelevant information, is a task humans perform subconsciously everyday. The ability to focus on details that are task-relevant, and abstract away those that are not, is considered cornerstone to human intelligence and information processing. Inspired by their ability to simplify problems by removing irrelevant details, researchers within the intelligent systems community have leveraged the power of abstractions to reduce the complexity of solving real-world problems in autonomous decision-making and control. However, despite their role in enabling autonomous agents to solve complex problems, the design of abstractions has been traditionally handled by system engineers, who provide heuristic, domain-specific knowledge that guides the construction of the reduced-order representations. For this reason, a growing interest in the development of frameworks that design task-relevant abstractions for autonomous agents has emerged, motivated largely by the central role of abstraction in intelligent systems.
To design task-relevant abstractions requires the preservation of relevant information through the process of compression. A formal treatment of these notions has been considered by information-theorists, who have developed a number of powerful frameworks for signal compression that rigorously capture the trade-off between relevant information retention and compression when encoding signals for transmission across capacity-limited communication channels. Of particular interest is the information-bottleneck (IB) framework, which formulates an optimization problem to design encoders that maximize compression while remaining maximally retentive regarding task-relevant information. In recent years, frameworks that employ IB-like approaches in order to design latent representations for autonomous systems have been developed, but with varying degrees of success, reproducibility, and theoretical guarantees. Motivated by these observations we, in this dissertation, develop frameworks that leverage ideas from information-theoretic signal compression to generate and design abstractions for autonomous systems. The frameworks allow for task-specific, multi-resolution, hierarchical tree abstractions to be obtained that are not provided to the system a priori, and instead emerge as a function of the agent's resource constraints.
In more detail, this dissertation contributes by drawing on the connection between hierarchical data structures and signal encoders to introduce an information-theoretic hierarchical tree-search problem which leverages the IB-principle to design multi-scale abstractions for autonomous systems that can be tailored to system resource-constraints. To solve our problem, we develop an algorithm, called Q-tree search, which employs a dynamic-programming-like pruning rule which we formally establish results in the optimal tree solution to our information-theoretic problem. Moreover, we show how the hard-constrained version of our problem may be realized as an integer linear program, thereby allowing multi-scale abstractions to be designed subject to hard system resource constraints, such as limited on-board memory. We discuss the connection between the two formulations, and establish a formal bridge between them by leveraging ideas from duality and relaxation theory. An algorithm for choosing the trade-off parameter in the soft-constrained, Q-tree search, problem as a function of the setting of the hard-constraint is proposed which leverages the structure of the problem and so-called tree phase-transitions to select the trade-off parameter by maximizing the dual function of the hard-constrained formulation. We then develop a framework that employs hierarchical abstraction, specifically those generated by Q-tree search, to reduce the computational complexity of planning in a principled manner that endows agents the ability to trade path-cost (quality) and environment information (resolution) in a rigorous fashion. A generalized formulation of the information-theoretic abstraction problem is then presented which considers the design of multi-scale hierarchical representations in the presence of multiple information sources. The generalized approach allows for both task-relevant and task-irrelevant information sources to be specified, and has connections with concepts from information-theoretic privacy. Importantly, the generalized method enables the creation of multi-resolution hierarchical abstractions of environments containing probabilistic semantic information, thereby allowing semantic-information driven abstractions to be generated that can be tailored to retain (and/or discard) desirable (undesirable) semantic classes (e.g., grass, asphalt, etc.). To solve the generalized problem, we develop the G-tree search algorithm and formally show that the proposed algorithm returns the optimal solution (i.e., multi-resolution tree). Finally, this dissertation contributes by developing a joint map-building and compression algorithm that simultaneously builds and compresses three-dimensional (3D) octree data structures containing (probabilistic) semantic information. The map-building and compression algorithm builds a Bayesian multi-class semantic octree from semantically-labeled point-cloud data which is subsequently compressed by a modified version of the G-tree search algorithm. We demonstrate the ability of our approach to compress large, semantically rich, outdoor environments built from real-world data, and show how the semantically-driven abstractions may be employed to create informed colored-graphs for semantic-planning.
Sponsor
Date Issued
2023-07-26
Extent
Resource Type
Text
Resource Subtype
Dissertation