Red Team/Blue Team Games with Contingency Planning and Adversarial Team Games

AbstractProject Title: Red Team/Blue Team Games with Contingency Planning and Adversarial Team GamesPI: Christian Kroer, Columbia UniversityProposal to Office of Naval Research# N00014-23-S-B001This proposal investigates game-theoretic models for red team/blue team force deployment problems,incorporating modern planning issues such as contingency planning, decentralized versus centralizedcontrol, and the importance of electronic surveillance technology such as drones. We propose to model thisproblem as a team game with sequential structure (since decisions are made over time) and imperfectinformation (since the exact state of the opposing team, or even some of the team's own units, is not known).In our team game model, the red and blue teams are each comprised of multiple agents, representingindividual units as well as a potential central planner. The two key difficulties in sequential team gamesare:1. Agents in a team do not know the precise state of the other agents on their team, e.g. the central plannercan only communicate approximate plans to units that have to operate with limited radio communication.In order to address this issue, teams must coordinate their actions, potentially by correlating their behaviorahead of time.2. The complex action space, sequential structure, and imperfect information implies that these models canbe very large, requiring the development of sophisticated optimization methods for computing optimalstrategies.In the first stage of our proposal, we explore a static game model where each individual unit is assigned alocation todefend or a route towards a target, and the goal of the central planner on the blue team is tooptimally assign these deployments, given that the red team will then perform optimal routing in order toreach their targets, while taking into account the (randomized) blue teams' strategy.In the second stage of the proposal, we will consider a sequential game model, where we allow individualunits to take actions, suchas whether and how often to synchronize information with the central planner,or to take decentralized decisions when there is little room for communication. We will model our problemsas adversarial team games and potentially also as Stackelberggames. For both settings, our work willleverage recent advances in the representation of the set of correlated strategies for agents on a team.Our model and results will also consider extensions to the inclusion of other assets such as various formsof sensors. Weexpect that the results we develop will extend to other patrol planning scenarios, such asthose encountered in wildlife protection games (known as green security games), drug traffickinginterdiction, and other geospatial patrolling problems.