Theory of Mind based Communication for Human Agent Cooperation

Abstract

For human agent cooperation, reasoning about the partner is necessary to enable an efficient interaction. To provide helpful information, it is important not only to account for environmental uncertainties or dangers but also to maintain a sophisticated understanding of each other's mental state, a theory of mind. Sharing every piece of information is not a good idea, as some may be irrelevant at time or already known, leading to distraction and annoyance. Instead, an agent will have to estimate the novelty and relevance of information for the receiver, to trade off the cost of communication against potential benefits. We propose the concept of theory of mind based communication as principled formulation to ground an agents cooperative communication on an understanding of the receiver's mental states to support her awareness and action selection. Therefore we formulate the problem of whether, when and what information to share as a sequential decision process with the human belief as central source of uncertainty. The agent's communication decision is obtained online during interaction by combining a second level Bayesian inference of human belief with planning under uncertainty, evaluating the influence of communication on the human belief and her future decisions. We discuss the resulting behavior on an illustrative communication scenario with different uncertain state aspects that an observing agent can communicate to the actor.

Using a Stochastic Agent Model to Optimize Performance in Divergent Interest Tacit Coordination Games

“…Overall, our findings can contribute to the development of agent models dealing with human–machine cooperation where collaboration is constrained by the cost of communication. In scenarios like these, the intelligent agent needs to trade off the cost of communication against its potential benefits [ 42 ]. Incorporating a behavioral model like ours into the agent model could guide the agent in deciding when it is worthwhile to communicate taking into consideration the associated cost.…”

MIRROR: Differentiable Deep Social Projection for Assistive Human-Robot Communication

“…Compared to standard black-box human models, MIRROR provides additional structure that can ease data requirements. Compared to handcrafted ToM approaches [2], MIRROR is able to handle high-dimensional observations. MIRROR is related to recent approaches that focus on capturing human traits, e.g., biases under risk and uncertainty [24] or action errors due to misunderstood environmental dynamics [25].…”

“…Prior work on planning communication methods in HRI typically rely on human models, which are typically handcrafted using prior knowledge (e.g., [1], [2]) or learned from collected human demonstrations (e.g., [3]). Unfortunately, handcrafted models do not easily scale to complex real-world environments with high-dimensional observations, and data-driven models typically require a large number of demonstrations to generalize well.…”

Assisted Perception: Optimizing Observations to Communicate State

“…Task-specific assistance via communication and visualization. Bühler and Weisswange (2020) assist users by modeling their internal beliefs and communicating observations that induce optimal actions, but require knowledge of the user's reward function at test time, assume a discrete state space, and do not learn a personalized model of the user's internal state estimation process. ASE does not assume knowledge of the task rewards at test time, can be applied to domains with high-dimensional, continuous observations like images, and interactively learns a user model.…”