Decentralized Autonomous Networks for Cooperative Estimation

Project Overview

While artificial intelligence, machine learning, and autonomous systems are seeing increasingly widespread use, solutions for commercial applications often are not designed for adversarial environments, contested communication links, and other challenging properties of national security applications. These situations require fully distributed data processing and autonomy algorithms which are robust to unreliable communication, falsified data, and node attrition. Solutions must also operate on edge computing devices of limited size, weight and power (SWaP), and scale to large network sizes. The Decentralized Autonomous Networks for Cooperative Estimation LDRD (DANCE, 20-SI-005) developed novel algorithms for decentralized data processing, interference, and autonomy required to leverage the advantages future autonomous sensor networks can provide in national security applications. The project produced novel theoretical results and algorithms for decentralized inference using both Bayesian techniques and robust deep neural network (DNN) ensembles, as well as adaptive collaborative autonomous behaviors which determine the future action of all systems in a network to improve future performance based on multi-agent reinforcement learning (MARL) and information theory. Uniquely, all algorithms in DANCE were developed and evaluated using realistic wireless communication models which accurately model packet loss and latency. In specific cases, we evaluated algorithms on low size, weight, and power hardware and developed novel methods for accelerating them with minimal sacrifices in accuracy. DANCE produced a holistic and generalizable approach to decentralized inference and autonomy with full uncertainty quantification which could be the foundation of future autonomous sensor networks in multiple applications.

Mission Impact

DANCE's approaches to decentralized sensing, processing, and decision-making are applicable to a wide range of applications of Laboratory interest, specifically when large network sizes and/or resilience is critical, and in dynamic or adversarial scenarios requiring fast adaptation in particular. Most notably, they enable networks of air, ground, sea, subsea, and/or space sensors to operate with only local communication links, replace vulnerable individual systems, and improve coverage and persistence due to the large network sizes they support. The onboard intelligence these algorithms provide can yield more effective sensor placement, more efficient or lower power operation, and coordination between different individual sensors and/or modalities, resulting in better situational awareness for several Department of Defense and Department of Energy missions. They are also applicable to critical infrastructure and command, control, and communication networks to increase robustness, response time, and resilience to adversarially-manipulated data in applications such as resilient power grid and microgrid operation.

Publications, Presentations, and Patents

Cadena, Jose, Priyadip Ray, Hao Chen, Braden Soper, Deepak Rajan, Anton Yen, and Ryan Goldhahn. 2021. "Stochastic Gradient-Based Distributed Bayesian Estimation in Cooperative Sensor Networks." IEEE Transactions on Signal Processing 69: 1713-24. https://doi.org/10.1109/TSP.2021.3058765.

Cadena, Jose, Priyadip Ray, and Ryan Goldhahn. 2021. "Decentralized Black-Box Variational Inference for Bayesian Learning on Sensor Networks." 2021 55th Asilomar Conference on Signals, Systems, and Computers, 150-55. Pacific Grove, CA, USA: IEEE. https://doi.org/10.1109/IEEECONF53345.2021.9723409.

Cadena, Jose, Priyadip Ray, Hao Chen, Braden Soper, Deepak Rajan, Anton Yen, and Ryan Goldhahn. 2021. "Decentralized Black-Box Variational Inference for Bayesian Estimation in Sensor Networks." IEEE Transactions on Signal and Information Processing over Networks. Under Review.

Chen, Cheng, Bhavya Kailkhura, Ryan Goldhahn, and Yi Zhou. 2021. "Certifiably-Robust Federated Adversarial Learning via Randomized Smoothing." 2021 IEEE 18th International Conference on Mobile Ad Hoc and Smart Systems (MASS), 173-79. Denver, CO,: IEEE. https://doi.org/10.1109/MASS52906.2021.00032.

Cheng, Hao, Kaidi Xu, Zhengang Li, Pu Zhao, Chenan Wang, Xue Lin, Bhavya Kailkhura, and Ryan Goldhahn. 2022. "More or Less (MoL): Defending against Multiple Perturbation Attacks on Deep Neural Networks through Model Ensemble and Compression." 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 645-55. Waikoloa, HI. USA: IEEE. https://doi.org/10.1109/WACVW54805.2022.00071.

Dawson, William A., Ruben Glatt, Edward Rusu, Braden C. Soper, and Ryan A. Goldhahn. "Hybrid Information-Driven Multi-Agent Reinforcement Learning." In Workshop on Challenges and Opportunities for Multi-Agent Reinforcement Learning (COMARL AAAI 2021), Palo Alto, CA. 2021. https://doi.org/10.48550/ARXIV.2102.01004.

Glatt, Ruben, Felipe Leno da Silva, Braden Soper, William A. Dawson, Edward Rusu, and Ryan A. Goldhahn. 2021. "Collaborative Energy Demand Response with Decentralized Actor and Centralized Critic." Proceedings of the 8th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation, 333-37. Coimbra Portugal: ACM. https://doi.org/10.1145/3486611.3488732.

Goldhahn, Ryan, Ray, Priyadip, Soper, Braden, Chen, Hao, and Rajan, Deepak. "Identification of a Characteristic of a Physical System Based on Collaborative Sensor Networks." US Patent #US20210241124A1, filed February 2, 2021.

Li, Qunwei, Bhavya Kailkhura, Ryan Goldhahn, Priyadip Ray, and Pramod K. Varshney. 2022. "Robust Decentralized Learning Using ADMM With Unreliable Agents." IEEE Transactions on Signal Processing 70: 2743-57. https://doi.org/10.1109/TSP.2022.3178655.

Rusu, Edward, and Ruben Glatt. 2021. "ABMARL: Connecting Agent-Based Simulations with Multi-Agent Reinforcement Learning." Journal of Open Source Software 6 (64): 3424. https://doi.org/10.21105/joss.03424.

Soper, Braden, Priyadip Ray, Hao Chen, Jose Cadena, and Ryan Goldhahn. 2022. "Bayesian Multiagent Active Sensing and Localization via Decentralized Posterior Sampling." In 2022 56th Asilomar Conference on Signals, Systems, and Computers. Pacific Grove, CA, USA: IEEE. Accepted for publication.