Contemporary Supervised Machine Learning (SML) and explainable AI (artificial intelligence) methods can be employed to both model and understand the decision making behavior of human actors within a multi-agent task setting. Here, we apply such modeling approach to capture the decision-making behavior of human actors playing a 3-player online herding game called “Desert Herding”. Of particular interest is whether the modeling approach can be employed to predict and understand the target switching strategies of human herders at variable prediction horizons and whether the explainable AI tool SHAP can be leveraged to identify the key informational variables (features) underlying the players’ target selection decisions.

Embodied virtual interactions can contribute to more immersive perspective taking experiences, which in turn can increase empathy and affiliation. This study sought to investigate these outcomes in the context of unconscious bias related to gender identity during interpersonal interactions. We conducted a simulated interview in virtual reality, in which participants embodied a transgender or cisgender avatar and interacted with a human-controlled agent (transgender woman). Preliminary results reveal differences between women and men in their experiences of empathy and emotional state when embodied as a transgender avatar.

Social motor coordination is an important mechanism responsible for creating shared understanding but can be a challenge for Autistic individuals. Social virtual reality (VR) provides an opportunity to create a safe and inclusive environment for which interactions can be augmented to promote social interactivity. Due to the bi-directional nature of social interaction and adaptation, we created a framework to explore social motor coordination with a virtual artificial agent which can exhibit human-like behaviors. In this experiment, we assessed the interactive behaviors of participants completing a collaborative problem-solving task with the agent using multidimensional cross-recurrence quantification analysis (mdCRQA). Our results show that participants who discovered novel solutions to the task exhibited greater coupling to the artificial agent regardless of participant characteristics. Future work will explore how social VR environments can be augmented to promote social coordination.

Recommender systems designed to augment human decision-making in multi-agent tasks need to not only recommend actions that align with the task goal, but which also maintain coordinative behaviors between agents. Further, if these systems are to be used for skill training, they need to impart implicit learning to its users. This work compared a recommender system trained using deep reinforcement learning to a heuristic-based system in recommending actions to human participants teaming with an artificial agent during a collaborative problem-solving task. In addition to evaluating task performance and learning, we also evaluate the extent to which the human action are congruent with the recommended actions.

A popular technique for evaluating envisioned AI agents without committing significant engineering resources to implement such an agent is the Wizard of Oz (WoZ) methodology. Thus far, many of the WoZ studies concerned with the design of conversational agents have been standalone systems for individual participants to interact with. The aim of this work was to determine whether a WoZ conversational agent could be embedded within an existing team-based collaborative virtual environment. In doing so, we uncovered a number of potential challenges in designing such a system which may be valuable for those implementing such a system in the future.

Working with others not only improves behavioral efficiency, but also facilitates learning. Such multiagent activity is fundamental to everyday life and, increasingly, virtual and robotic agents are finding a place in these contexts. The effectiveness of human-machine interaction (HMI), however, relies on artificial systems being able to anticipate their partner and select actions that not only lead to achieving the shared goal, but does so efficiently. Here, a multiagent "shepherding" task was used to study coordination and behavior-switching during HMI. The task required the coordinated control of a complex environment, where a non-obvious solution leads to near-optimal task performance. Previous research has demonstrated that a virtual agent, with knowledge of the optimal solution, can effectively steer novices to discover the optimal task behavior [3]. Conversely, results here demonstrate that when completing the task with a virtual avatar incapable of producing this behavior, a subset of novices still discovered and enforced this optimal behavior in the virtual avatar by modulating the sheep-herd's dynamics. These results provide evidence that learning efficient solutions may result from interaction patterns early in the interaction, which may be exploited by adaptive artificial-agents in HMI contexts to facilitate skill acquisition.

Humans often engage in tasks that require or are made more efficient by coordinating with other humans. The coordination involved in these tasks can be understood in terms of the behavioral and affordance dynamics of socially embedded agents engaged in joint action activities. Behavioral dynamics provide mathematical (differential equation) models of human behavior and interaction and affordance dynamics identify and model the ways that an agent's action capabilities evolve over time. Taken together, models of human joint-action based on these approaches may provide a basis for developing robust, natural, and easy to engage artificial agents. In this paper we introduce behavioral and affordance dynamics models of human joint action in a pick-and-place task. Based on these models we provide a proof of concept pick-and-place artificial agent and implement the agent in a 3D virtual environment to interact with human co-actors.

Models of robust human-human coordination can guide the design of adaptive and responsive human-robot systems. Here we test an artificial agent that embodies low- dimensional nonlinear dynamic equations derived from human behavior while completing a two-agent herding task, where the goal is to contain reactive spheres to the center of a target region. The model was able to complete the task alongside human novices in a virtual version of the experimental setup used in Nalepka and colleagues (submitted). Not only did the model lead participants to successful performance, but also 12 out of 18 participants reported that they believed their partner was a human participant in another room. The model was therefore able to capture the complex social behavior that defined robust task success in terms of lower dimensional dynamical equations that characterizes the emergent behavioral dynamics of embedded multiagent behavior.

Interpersonal or multiagent coordination is a common part of everyday human activity. Identifying the dynamic processes that shape and constrain the complex, time-evolving patterns of multiagent be ...