Karaoguz, Cem: Learning of information gathering in modular intelligent systems. 2012
Inhalt
- Introduction
- Scope of the Thesis
- Biological Background
- Modularity in Cognition
- Attention and eye movements
- Stimulus-driven and task-driven attention
- Working memory
- Reward-based learning and attention
- Related Work
- Overview
- Modular Systems Approach
- Modularity in Artificial Systems
- Adopted Modular System Model
- Formal Definition of the System
- Functional Decomposition of the System
- Benefits of Modularity in Handling System Complexity
- Summary
- Learning in Large Scale Modular Systems
- Reinforcement Learning
- Markov Decision Process (MDP)
- Optimizing Value Functions
- Modular Reinforcement Learning
- Partial Observability
- Module Management Learning
- Module Management Framework
- On-line Learning of the Parameters
- Regulation of the Learning Process
- Empirical Observations of MML
- Summary
- Active Vision Application
- Why Active Vision?
- System Architecture for Active Vision
- Outline of Experiments
- Results and Discussion
- Overall System Performance
- Support for the Reaching Task
- Support for the Human Interaction Task
- Emerging Gazing Behaviors
- Individual Contributions of Learning Mechanisms
- Adaptation to Changes
- Analysis of the MML Mechanism
- Summary
- Autonomous Navigation Application
- System Architecture for Autonomous Navigation
- Outline of Experiments
- Results
- Benefits of Modular Systems Approach
- Overall System Performance with Constrained Resources
- Learned Information Acquisition Behavior
- Support for Policy Learning
- Analysis of the MML Mechanism
- Summary
- Conclusions