Robust and Adversarial Model Learning, Optimization, and Human-Aligned Decision Support
Abstract
AbstractN00014-22-S-B001Program Office Code: 311 [Krebs, William]Robust and Adversarial Model Learning, Optimization, and Human-Alig,ned Decision SupportProf. Matthew Gombolay (PI) ? Georgia Institute of TechnologyIn this proposal, we will develop new computational, methods, open-source implementations, and human-machine interfaces to (Aim 1) to infer models of adversarial, multi-agent team beha,vior, (Aim 2) construct new robust, adversarial optimization mechanisms for reinforcement learning to allocate resources to accompli,sh a mission while being robust to adversarial attacks, and (Aim 3), incorporate commander?s intent to inform planning.Aim 1) Learni,ng Models of Adversaries ? In our first aim, We propose new approaches for learning modeling adversary actions in adversarial games,with heterogeneous graph attention networks. Specifically, we propose a parallelized approach for learning filtering and prediction,to promote information sharing across the two modules during training, and mitigate error propagation from filtering to prediction,,if connected sequentially. While most prior works in adversarial opponent modeling have been restricted to small grid worlds, we see,k to create a large-scale, novel domain for modeling adversarial target tracking in real- world scenarios with a variable number of,tracker agents and diverse terrains that impact the visibility of those tracker agents.Aim 2) Optimizing Robust Plans ? In our next, aim, we develop model-based planning approaches for allocated resources (e.g., naval platforms) in an adversarial environment. Spec,ifically, we consider the game of multi-agent task assignment and routing and the meta-game of selecting agents? initial locations.,We propose leveraging heterogeneous graph attention networks, which can generalize to unseen problems consisting of an arbitrary num,ber of agents with varying parameters. We also propose extending prior work in this initialization meta-game to be adversarial, such, that one team is performing the task assignment and routing problem, and the other team is attempting to impede the assignment and,routing agent. Each team may have only partial observations of their adversary.Aim 3) Supporting Commander?s Intent ? Finally, we s,eek to develop the human- machine interfaces that incorporate commander?s intent via natural language to develop human-aligned plans,. While significant work has been conducted to instruct a planner to solve a task, via language or demonstrations, prior work lacks,a focus on building planners which can operate within the parameters specified by a team. Worse yet, there is a dearth of research p,ertaining to enabling humans to provide their specifications through unstructured, naturalistic language. In this aim, we will devel,op new technical methods for natural language processing to understand a human?s desired specifications for a plan and develop new h,uman-machine interfaces for recommending a right set of plans to enable the human to find a plan that aligns with expectations.Appro,ved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 07, 2022
- Source ID
- N000142212834
Entities
People
- Matthew Gombolay
Organizations
- Georgia Tech Research Corporation
- Office of Naval Research
- United States Navy