You are here

Deep Inference using Strategy Clustering over Embedded Representations (DISCERN)

Award Information
Agency: Department of Defense
Branch: Navy
Contract: N68335-20-C-0584
Agency Tracking Number: N201-077-0291
Amount: $139,957.00
Phase: Phase I
Program: SBIR
Solicitation Topic Code: N201-077
Solicitation Number: 20.1
Solicitation Year: 2020
Award Year: 2020
Award Start Date (Proposal Award Date): 2020-06-08
Award End Date (Contract End Date): 2020-12-08
Small Business Information
625 Mount Auburn Street
Cambridge, MA 02138-4555
United States
DUNS: 115243701
HUBZone Owned: No
Woman Owned: No
Socially and Economically Disadvantaged: No
Principal Investigator
 Jeff Druce
 (617) 491-3474
Business Contact
 Erica Hartnett
Phone: (617) 491-3474
Research Institution

Identifying courses of action (COAs) in strategic engagements is complex and time-intensive, and subject to fluctuating mission state conditions. Although AI-enabled game-playing algorithms have proved proficient at identifying both optimal and diverse strategies, their outputs are often inexplicable, and therefore cannot be leveraged for decision support. This is a critical limitation to deploying AI for decision aid in this domain, as Warfighters need to be able to assess the optimality and underlying objectives embedded in multi-objective AI-recommended COAs prior to executing. What is needed is an AI-enabled yet understandable decision aid that can support Warfighters in exploring the space of COAs available under different mission states. To address this issue, Charles River Analytics is pleased to propose Deep Inference for Strategy Clustering using Embedded Representations (DISCERN), a generative, hierarchical deep reinforcement learning framework with an inference engine that identifies, labels, and ranks sub-decision track clusters. These sub-decision tracks will represent subsets of decisions which produce a human-understandable objective, or micro-task, in a multi-objective scenario. DISCERN’s explainable sub-decision track recommendations are intrinsically tied to these corresponding objectives to generate detailed, COA recommendations expected to lead to achievement of a specific objective. To support Warfighters with these recommendations, under the DISCERN effort, we will: (1) develop a generative framework for associating low level action sequences with corresponding micro-task objectives, (2) develop an inversion model to infer micro-tasks from observed action sequences, and (3) identify, cluster, and rank sub-decision tracks from observed mission states and display our analysis in an intuitive decision support tool for Warfighters. Our approach will be demonstrated in the Starcraft2 simulation environment.

* Information listed above is at the time of submission. *

US Flag An Official Website of the United States Government