Human-guided Online Learning for Long-term Autonomy with Active Self Evaluation
Abstract
Developing unmanned systems that demonstrate long-term autonomy in highly uncertain and poorly understood environments is critical for future Army operations. Harnessing intelligence from humans can fundamentally revolutionize the capability of unmanned systems, especially for long-term autonomy, because humans have shown unique talents in making long-term smart decisions in the open world. Recent advances in long-term autonomy mainly focus on static environments in, e.g., indoor manufacturing and logistics. These advances, however, can hardly be applied in uncertain and dynamic environments due to the challenges to enable some key long-term autonomy features, such as, learn from and exceed humans, explain long-term decision intelligence, and evaluate decision effectiveness online. The goals of the project are to overcome theoretical and algorithmic challenges via developing a human-guided online learning approach that integrates a newly designed stereo reinforcement learning and explainable multi-objective learning with human-guided learning to enable self online evaluation with need-based human guidance. Specifically, the project will focus on three essential thrusts: (1) Stereo Reinforcement Learning: simultaneously learn rewards in humansÕ decisions and generate control policies that outperform humansÕ decisions; (2) Explainable Multi-objective Learning: construct new algorithms that can learn the tradeoff principles among different objectives in humansÕ long-term decision making; and, (3) Active Online Evaluation: create an online evaluation method that judges the decision effectiveness to determine if any further human guidance is needed. The success of the project is expected to fulfill current needs at the U.S. Army by providing data-efficient, human-friendly, and self-evaluation unmanned systems. The novelty of this project is the synthesis of approaches from cognitive, artificial intelligence, computational, learning, decision/control sciences to create a new human-guided learning approach that offers efficiency, interpretability, self-awareness, and logic-thinking.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110103
Entities
People
- Yongcan Cao
Organizations
- Army Contracting Command
- United States Army
- University of Texas at San Antonio