Distributed Control and Information Fusion Over Communication Networks
Abstract
This project focused on distributed control and information fusion/learning over communication systems. The first set of problems considered were related to distributed adaptive control. Namely, a first such decentralized learning algorithm for multi-armed bandit models was developed that achieved poly-logarithmic regret. Later another algorithm was given that achieve log regret. The second set of problems was related to decentralized control design for LQG systems with asymmetric one-step delayed information sharing between two controllers. The optimal decentralized controller was computed. A case with partial output feedback was also solved. The third problem solved was development of empirical value iteration, a new method for approximate dynamic programming. Convergence of such an algorithm to the optimal policy was established by introducing new notions of probabilistic fixed points of random operators. Numerically, this new class of methods is found to perform better than reinforcement learning methods.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 17, 2013
- Accession Number
- ADA584502
Entities
People
- Rahul Jain
Organizations
- University of Southern California