MineApprentice
Navy SBIR FY2018.1


Sol No.: Navy SBIR FY2018.1
Topic No.: N181-079
Topic Title: MineApprentice
Proposal No.: N181-079-0251
Firm: Soar Technology, Inc.
3600 Green Court
Suite 600
Ann Arbor, Michigan 48105
Contact: Christopher MacLellan
Phone: (734) 627-8081
Web Site: http://www.soartech.com
Abstract: Automated planners have been adopted as key decision support tools for mission planning in a wide range of domains, such as maritime Mine CounterMeasures (MCM). However, the models that power these planners, such as models of capabilities, tactics, environment, and goals, must currently be hand-codedÉ?"a tedious and slow process that makes model maintenance impractical for highly dynamic domains where models must be constantly updated. Further, the hand-coding approach does not scale to scenarios with increasing numbers of heterogeneous unmanned assets, such as unmanned underwater vehicles, where their sensing and autonomous capabilities can vary widely and change far more rapidly than larger manned systems. In these scenarios, MCM operators cannot rely on automated planners to have adequate knowledge of assets and how best to tactically employ them because the rate at which this knowledge can be hand-coded is insufficient to keep up with the tempo of change. These limitations ultimately result in suboptimal mission performance and underutilization of unmanned assets. To overcome these challenges, the SoarTech team proposes MineApprentice, a machine learning system that incrementally and continually acquires both asset performance models and tactical planning knowledge from interaction with operators as they use existing MCM planning software (e.g., MEDAL-EA/MineNet Tactical).
Benefits: MineApprentice leverages two innovative SoarTech approaches: (1) probabilistic concept formation, which incrementally builds a probabilistic taxonomy of previously used assets and how they performed under previous environmental conditions in order to predict performance for new assets and/or new environmental conditions, and (2) interactive task learning, which enables operators to update a systemÉ?Ts planning knowledge by critiquing system generated plans and by providing examples of desirable planning behavior. These approaches will enable MineApprentice to improve coverage and performance of MCM mission planners and substantially decreases operator planning timeÉ?"ultimately leading to increased adoption of automated planners and improvement in overall measures of mission effectiveness (decreased time and risk).

Return