Charles River Analytics was awarded a $140,000 contract from the US Navy to develop Deep Inference using Strategy Clustering over Embedded Representations (DISCERN). DISCERN is an easy-to-understand, AI-enabled decision aid that helps humans better understand reinforcement learning agents.
In reinforcement learning, a programmer trains an AI agent to respond to the surrounding environment. These agents are often hard for humans to understand—they operate based on low-level behavior sequences that require humans to reason about, or imagine, the higher-level objectives guiding their behavior.
DISCERN is a hierarchical, deep reinforcement learning framework coupled with an inference model that identifies these higher-level objectives and associated confidence metrics. To create this framework, we employ Neural Program Policies, a novel class of deep RL policy models developed at Charles River Analytics, to train agents with built-in hierarchical structure to play StarCraft II.
Charles River Analytics researchers think that associating agent behavior with the strategic choices made by an AI agent at various levels of abstraction will provide a powerful explanatory tool. The insights from this tool can then be applied to other applications, such as strategy development and AI-assisted decision-making.
This material is based upon work supported by the Office of Naval Research under Contract No. N68335-20-C-0584. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Office of Naval Research.