CONTACT: Collaborative Neural-agents providing Team Assistance in Cognitive Tasks
Abstract
Approved for public release. -Tasks that require weighing evidence of competing narratives and evidence and coming to decisions on downstream actions are by and large relegated to humans rather than AI. AI assistance in such tasks remains an open challenge, as the human judgement, experience and training for experts to perform these tasks cannot be easily codified and gathered at scale for training AI agents. For example, naval teams that plan deployments for staging of personnel and assets dynamically incorporate multiple streams of information from intelligence briefs regarding environment conditions. In these and other settings, there is an opportunity to employ learning of context-dependent workflows and embed intelligent, collaborative agents in the workflow to proactively highlight information, conduct data synthesis, and recommend task steps, thereby reducing variability, and improving consistency and quality across experts performing sequential decision-making tasks. --This project will leverage the increasing sophistication of deep RL-trained agents performing complex sequential decision-making tasks, and investigate settings in which the action/decision chains of the AI agent cannot be parsed and separated from human decision-making steps. We will focus on settings in which a person is required to bring their judgement and expertise to determine when and how to rely on the decisions and inputs of the AI agent over time. We will build on our prior research and investigate settings in which the agent will collaborate to support the complex task of dynamically updating a disaster response deployment plan in the face of arrival of new intelligence briefs, in the presence of asymmetric information among human and agent team members. Agent collaboration in this setting is challenging because human teams often form plans by achieving consensus. In addition, machine support of team planning requires computationally tractable models to infer shared understanding. Finally, machine participation in real world team planning is especially challenging when the machine assumes its models are correct and complete. This project will build on PI Shah s prior research, which introduced computationally tractable models for monitoring the team s planning discussion and inferring shared understanding, with the intent to develop a collaborative agent that can participate online in human team planning to strengthen the quality of the plans produced. --We envision a collaborative agent that analyzes communications among a team of human planners to infer and maintain a belief over the team s evolving plan. The agent will also analyze incoming intelligence briefs and cooperate - helping to improve human-directed outcomes - by performing concept classification tasks of natural language text of the briefs - in which domain experts (non-ML experts) craft and refine concepts important to eparticular tasks and settings. --The goal is to develop and prove-in RL agents that learn to proactively communicate relevant information from intelligence briefs at the right time, and propose high-level strategies to the human planners for how to update the plan. PI Shah s prior research has demonstrated that incorporation of human planners high level strategies using preferences, which are composed of soft goals and soft trajectory constraints, improved the quality of automated plans by without substantially increasing plan computation time. These plans also achieved greater action similarity to users# manually-generated plans, as compared to plans that were generated without the use of user-provided strategies. The goal in this project is to endow the collaborative AI agent with the same ability to make salient the relevant information at the right time and recommend high level strategies for refining the human team s planning in a productive and unobtrusive manner.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 13, 2023
- Source ID
- N000142312883
Entities
People
- Julie Shah
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy