Applied Reinforcement Learning Wargaming with Parallelism, Cloud Integration, and AI Uncertainty
Abstract
The recent rapid growth in machine learning and artificial intelligence has created a growing opportunity to improve Department of Defense (DOD) wargaming. This study aims to leverage modern frameworks, algorithms, and cloud hardware to improve the DOD's wargaming capabilities, with a specific focus on reducing training times, increasing deployment flexibility, and demonstrating how a trained neural network can provide a measure of certainty for recommended actions. This work utilized an open-source framework for parallelization to train and deploy a neural network to the Azure cloud platform. To measure the certainty of a trained networks move choices, Bayesian variational inference techniques were employed. The application of the open-source framework resulted in a more than tenfold reduction in training times without any loss in performance. Additionally, deploying trained models to the Azure cloud platform effectively mitigated infrastructure constraints and Bayesian methods were successful in providing a measure of trained models certainty. The DOD could leverage these advancements in machine learning and cloud computing to significantly enhance future wargaming efforts.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2023
- Accession Number
- AD1213264
Entities
People
- Matthew G. Finley
Organizations
- Naval Postgraduate School