Mobius Logic Inc.

Our project will address Capability 3: Mission Command and Control. In many cases, the real-time detection and anticipation of other actors’ (opponents' or otherwise) intentions is critical to the success of an actor in multi-actor, goal-directed situations. Automating this detection is particularly important in the world of semi-autonomous systems. In practice, an opponent's intent can often be mapped to a series of decision points, where they may take actions given their perceived likelihood of success. Games like Texas Hold’ Em Poker provide a good testbed for using these decision points to automatically detect an opponent’s intent mid-game, and adjust our own strategy accordingly. We will begin by creating LLM agents with different poker player “personalities” (e.g. a reckless agent that often chooses to go all-in, a cautious agent that folds when they are unlikely to win, etc.). We will then collect observations from these agents and test clustering algorithms like BIRCH, Spectral, and Ward to differentiate the personalities in the observation space. Success from here involves finding the points in the game where the personalities can be differentiated and exploiting them online to force an adversary to reveal their personality. E.g. one decision point may occur immediately after we choose to raise following the flop. Based on our observation classification, we may know that a reckless opponent would choose to go all in here, whereas a cautious agent may choose to call. By choosing to raise at this point, we compel our adversary to choose an action consistent with only one of our personality paradigms, thus giving us insight into its intentions. The methodology we develop and test here will help form the groundwork for decision-making procedure that lets us determine and drive opponent actions in real time. It has many applications to Command and Control, including online intent determination, inverse RL, and threat classification. In addition, it can be used offline to enhance Agent Training and identify gaps in the known space of possible agent behaviors.