Situation Awareness and Adaptive Decision-Making in Autonomous Systems via Symbolic Learning

Open Access
Author:
Jin, Xin
Graduate Program:
Mechanical Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
September 04, 2012
Committee Members:
  • Asok Ray, Dissertation Advisor
  • Jeffrey Scott Mayer, Committee Member
  • Shashi Phoha, Committee Member
  • Alok Sinha, Committee Member
  • Thomas Wettergren, Special Member
Keywords:
  • situation awareness
  • decision making
  • autonomous system
  • machine learning
  • autonomous navigation
  • sensor fusion
Abstract:
Autonomous systems are becoming prevalent in the human society as they play a very important role in the applications that need to function in uncertain environments. Such systems usually require integration of several functions including perception and decision-making for operation under incomplete or incorrect a priori information. In this context, situation awareness is the perception of the environment for extraction of pertinent information, and adaptive decision-making is the adaptation to the uncertain environment. This dissertation addresses two main issues: situation awareness and adaptive decision-making in autonomous systems by symbolic learning. Situation awareness involves the recently developed framework of Symbolic Dynamic Filtering (SDF) and is discussed in the context of supervised learning. In the SDF framework, sensor observations are discretized temporally and spatially to generate blocks of symbols and patterns are generated from these symbol blocks for pattern classification. Research issues such as feature extraction from two-dimensional domain of wavelet-transformed sensor time series, optimization of the symbol generation process, and fusion of heterogeneous sensor information are addressed in this dissertation. The proposed technology is validated by various (experimental and simulated) case studies that include behavior recognition in mobile robots, anomaly detection in nuclear power plants, and personnel detection using multi-modal sensors. Adaptive decision-making is crucial to ensure effective and efficient operation of autonomous systems in uncertain environments, and is discussed in the context of path planning and autonomous navigation. The search space of the autonomous system is constructed as a symbolic grid map, where the two-dimensional space is partitioned into grids and an alphabet of symbols is assigned to each grid cell to represent the environmental information. Formulation of the search space is then extended in the multi-resolution sense to enable adaptive decision-making based on the available spatio-temporal information. A navigation algorithm based on the multi-resolution formulation is then proposed for autonomous navigation in uncertain environments. The proposed algorithm is validated on various case studies such as complete coverage of complex and static environments, and simulation of oil spill cleaning in dynamic and uncertain environments.