NG: Seed Grant in RL, AL, ML

Faculty

John S. Baras

Funding Agency

Northop Grumman, through the UMD Vice President for Research

Year

2023

Descriptions

This $50K seed grant supports a graduate student dedicated to the work below.

Decision-making speed, adaptivity, flexibility, robustness, reliability, and the ability to handle unexpected changes and events, have been key concepts in Mission Planning and Operation. The US DoD missions of interest are dynamic; environment, mission goals and plans, mission participants, can change dynamically and unexpectedly. Substantial progress has been achieved in the associated challenges by the integration of methods from dynamical systems, optimization, operations research, decision theory, logic and constraint programming.

A most recent trend, necessitated by the increasing complexity of missions and associated requirements, is the infusion of machine learning (ML), reinforcement learning (RL) and artificial intelligence (AI) in mission planning and operation. However, progress in the theoretical foundations and practical real-time applications of this infusion for mission planning-operation-evaluation has been limited.

There are several reasons. First, lack of theory for integrating model-based and data-based (i.e. ML and AI) approaches. Second, infusion has been an “add-on”; not considered from start in mission planning and operation. Third, ML and AI components require large amounts of data for training that are not available for mission related applications. Fourth, works in this trend fail to address time and resource constraints essential for practice. Fifth, they do not consider risk (in various forms) as a fundamental concept and the associated tradeoff with performance (utility).

This research addresses problems under the AI/ML topic of the solicitation for the 2023 Northrop Grumman and UMD Seed Grant Program. It specifically aligns with the following subtopics: Robust, understandable, highly trusted mathematical algorithmic approaches; and capabilities development employing AI/ML technologies where high-level command and control functionality can be realized without the need of a human in the control cycle.

The research will develop novel concepts, methods, algorithms and systems for robust interactive ML and AI, that progressively learn repeated uses and significantly assist humans in mission planning and operation. It is anchored on ideas, concepts, methods, from earlier research by the principal investigator on: progressive universal ML, integration of knowledge representation and reasoning (KRR) with ML, robust interactive ML and AI via risk-sensitive decision making, time and delay concepts and challenges.