Charles River Analytics has secured $3 million award to develop tools that can help the U.S. Navy and Marine Corps better understand deep reinforcement learning, a machine learning technique restrained by obstacles associated with unclear responses and the potential to make wrong conclusions.
The company said Tuesday its Explainability and Terrain Reasoning for Autonomy or EXTRA effort will use artificial intelligence research and best practices in human-machine interface design to create tools capable of boosting the clarity of DRL agents.
“For human-machine teams to be able to engage in novel military tactics together, people need explanations from the AI that they can understand and trust,” said Jeff Druce, senior scientist at Charles River Analytics and co-principal investigator of the EXTRA effort.
Druce and fellow Charles River scientist James Niehaus serve as co-principals for the effort, which will train DRL agents in environments that embody the operational complexity witnessed by Sailors and Marines.
Their team will also produce tools that explain DRL agent decisions in a way understandable to humans.