Deep lifelong reinforcement learning for resilient control and coordination
Abstract
The objective of this effort was to develop deep lifelong learning methods that can successfully handle sequential decision making in complex dynamic environments, focusing on multi-agent intelligence, surveillance, and reconnaissance (ISR) scenarios. We developed a novel architecture for deep convolutional neural networks that supports lifelong learning via deconvolutional factorization (DF-CNN), explored a combination of policy distillation via Distral and Sobolev training, and developed a hybrid controller for applying deep learning to ISR agents. Our approaches were evaluated on standard benchmark deep learning datasets, the DOOM environment, and on ISR scenarios in the ATE3 simulation environment.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2019
- Accession Number
- AD1074259
Entities
People
- Eric Eaton
- James Stokes
- Seungwon Lee
Organizations
- University of Pennsylvania