Long Range Broad Agency Announcement (BAA) for Navy and Marine Corps Science & Technology
Abstract
Approved for Public ReleaseThe objective of this proposal is to develop algorithms and supporting theory for distilling high-level task knowledge from large-scale generative language models and for integrating such knowledge into downstream sequential decision-making algorithms that are based on formal synthesis and reinforcement learning. The resulting algorithms will enhance capabilities for the design of autonomy strategies by alleviating currently standard yet limiting assumptions, e.g., the existence of precise and complete knowledge of the atomic propositions and features necessary for a given task and by automating several steps in the design flow for autonomous systems. Additionally, we propose to investigate how the outcomes of downstream decision-making can be utilized to interrogate the distilled knowledge and to refine the prompts to language models accordingly.While large generative language models have emerged as a new potential tool that can be used in the design and operation of autonomous systems, the current understandingof their capabilities and limitations is a significant barrier to their adoption. The proposed effort takes an initial step toward realizing the viewpoint of utilizing their capabilities while safeguarding against their limitations. It is structured into three thrusts. Thrust I focuses on distilling task knowledge, as discussed above, in a form that can be incorporated into sequential decision-making. It then integrates such knowledge into two complementing approaches for sequential decision-making: Planning based on formal verification and synthesis (Thrust II) and task-guided reinforcement learning (Thrust III). Each thrust will also investigate using multi-modal generative models jointly for language and image. In addition to outputting strategies for sequential decision-makingwhen the intended strategies exist, we expect that Thrusts II and III will produce diagnostic information that can be used to interrogate the outcomes of Thrust I in order to revise the prompts to generative language models and to refine the resulting automaton-based knowledge representations.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 24, 2024
- Source ID
- N000142412097
Entities
People
- Ufuk Topcu
Organizations
- Office of Naval Research
- United States Navy
- University of Texas at Austin