Statistical Evaluation of Multi-Agent Reinforcement Learning Models Under Different Versions of TensorFlow
Abstract
The Agents Leveraging Learning for Intelligent Engagement with Soldiers (ALLIES) team has been using TensorFlow, a machine learning software library, to train and evaluate agents in tasks such as a continuous 2D version of the Predator-Prey Pursuit game. These tasks have provided a practical, dynamic research environment for studying cooperation and competition in agent-agent and human-agent teams, but staying compatible with research partners requires that we update to a new release of TensorFlow. To maintain continuity in our research, we evaluated predator-prey data created under the original and the new release of TensorFlow and confirmed that this transition does not affect training and evaluation.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 22, 2021
- Accession Number
- AD1125841
Entities
People
- Derrik Asher
- Erin Zaroukian
- Rolando Fernandez
Organizations
- United States Army