With the prevalence of advanced technologies in society, we can easily envision a future of adaptive and individualized systems that function with an individual’s capabilities and limitations to achieve greater human-system performance. This individualized human-technology approach is expected to enable greater variety in human behavior, while having the ability to maintain consistent, robust outcomes when viewing the human-technology behavior as a system. However, when considering multiple agents and multiple humans, much work still needs to be done to fully realize the envisioned future of human-agent teaming. Here we draw on key findings from research into individualization technologies, human teams, intelligent agent teams, and mixed agent teams to identify a foundation for future research questions on how individualized technologies can enhance human-agent teamwork.
Recently, we have seen unparalleled advancements in sensor and analysis technologies that provide new insights into different facets of human psychology, physiology, behavior, and performance. For example, advances in neuroscientific tools have revealed fascinating discoveries on how differences in brain structure and function are associated with precise human behaviors (Telesford et al. 2016; Garcia et al. 2017). Advances in social and environmental sensing tools have provided unprecedented insights into patterns of gross human social behaviors (Kalia et al. 2017), while advances in biochemical or fluid sensing (i.e., blood, sweat, and tears) are providing unique insights into the continuous dynamics of internal human states and traits. More generally, advances in wearable devices have enabled the tracking of a wide range of factors including activity, sleep patterns, and various physiological parameters (Bonato 2010). These advances can be coupled with novel computational methods to infer motivations, predict behavior, and reason about the environment and the agents acting in it.
Research on human teams started out very disparate; however, a growing consensus amongst experts in the field suggests the importance of a core set of team inputs, as well as attitudinal, behavioral, and cognitive emergent states and processes that impact performance-related outcomes (Marks et al. 2001; Ilgen et al. 2005; Salas et al. 2005; Burke et al. 2006; Salas et al. 2009). Team cognitive processes and states relate to shared cognitive activities such as shared situation awareness (SSA), shared mental models, transactive memory, and macrocognition (Cooke et al. 2007), while affective/motivational team processes and states include concepts such as team cohesion, collective efficacy, and intragroup conflict. Behavioral processes and states represent what teams actually do, or their actions, to produce team performance outcomes, such as communicate, coordinate, and adapt (Kozlowski et al. 2015). From this work, emerges a core set of actionable properties of teams that can be targeted for performance enhancements. However, questions remain regarding how these states and processes translate in human-agent teams.
Additionally, there is a growing body of literature on team composition and team assembly, which examines the influence of individual factors, relational and multimodal networks, and ecosystems of teams on group outcomes such as satisfaction and performance. While much work has linked semistable, individual compositional elements of teams (e.g., personality, cognitive ability and styles, demographics, knowledge and ability) to group-level processes and outcomes (see Cooke 2015 for review), less is known regarding the influence of individual characteristics in dynamic, long-term team contexts, where individual attributes may vary within the group and within individuals over time. Harnessing this within-group and within-individual variability in a team context requires a careful methodological approach to avoid risking loss of data richness when aggregating attribute data across individuals over time and across groups. Currently, little is known about how dynamic individual variability contributes to team performance, primarily because measuring continuously variable individual attributes is difficult and incorporating these variable attributes into models of group interaction dynamics has been methodologically infeasible until very recently (Schecter and Contractor 2016). In addition to recent advances in networked-based models that enable examination of more complex dynamics, incorporating individual variability into group-based performance measurements, dynamic measurement of team emergent properties and group performance has also advanced. Kozlowski et al. (2016) provided a research paradigm for examining the multilevel dynamics of emergence using computational modeling, simulation, and experimentation approaches. Simultaneously, many researchers have been making great progress in developing continuous, unobtrusive measures of these dynamic team processes, using sensor- and systems-based data (e.g., (Olguın et al. 2009; Baard et al. 2012; Rosen, Wildman, Salas, & Rayne 2012; Kozlowski et al. 2013; Orvis et al. 2013; Duchon et al. 2014)). While research examining the relationships between individual dynamics and team outcomes in relation to group composition and assembly remains scarce, these recent advancements in computational approaches and measurement of individual and team dynamics provide an opportunity to develop individualized approaches targeting team states, processes, and performance over time.
As opposed to the human-only teams research, which looks across this broad set of states and processes to enhance team performance, research on intelligent-agent teamwork tends to focus on coordination in teams of intelligent agents posed as problems of task allocation (Gerkey and Matarić 2004). Given some fixed team task and a team of individual agents with specified capabilities, how can the taskwork be optimally distributed? In human-only teams, this relates to division of labor, role clarity, and explicit coordination (Van de Ven et al. 1976; Kogut and Zander 1996). Although a difficult computational problem, both exact and heuristic methods already exist to solve these problems. Another potentially more difficult aspect of coordination in intelligent-agent teams is the ability to develop strategies and policies for teamwork in real-time to respond to the unique environment and team makeup instead of relying on preplanned or rule-based strategies (Stone et al. 2010). However, recent breakthroughs in deep reinforcement learning have led to the ability to learn complex cooperative behaviors in multiagent systems in an end-to-end framework (Foerster et al. 2016; Sukhbaatar et al. 2016). In addition, significant advances in terms of distributed optimization have laid the theoretical groundwork for distributed cooperation in teams of intelligent agents (Nedic et al. 2010). For evidence of the rate of improvement in teams of intelligent agents, the RoboCup competition provides one such example of cooperation of fully autonomous systems in a complex environment.
The question of how to merge advances in human team-focused research with advances in the intelligent agent teams research remains a challenge for enhancing teamwork in human-agent teams; however, considerable research has explored coordination in mixed-agent teams. The extant literature has most commonly offered substitution-based function allocation to balance exclusive control or decision authority between humans and autonomous systems (Sheridan 2000; Dekker and Woods 2002). Some function allocation concepts have considered task type and the level of autonomy (Parasuraman et al. 2000) alongside typical “man-is-better-at”-“machine-is-better-at” roles (Fitts 1951). Such function allocation concepts have been instantiated in a number of different control frameworks, the most widely recognized of which is supervisory control (Sheridan 1992), which can be implemented in a variety of ways ranging from autonomous waypoint navigation to shared control schemes in which both the human and the autonomous system provide control inputs with different relative contributions (e.g., Crandall and Goodrich 2002).
Call for Comments:
- What other advancements will enable individualizable and adaptive team-enhancement technologies?
- What barriers will prevent individualizable and adaptive team-enhancement technologies?
- Additional related comments