INTELLIGENT COACHING AGENTS FOR ENHANCING HELPING BEHAVIOR IN HUMAN TEAMWORK
Teamwork is the joint work of individuals who act together productively. Recent technological advancement, global competition and world events made teamwork vitally important to the success of many organizations in both military and civilian sectors. Team training is arguably one of the most intensively studied topics for cognitive science researchers in the past decades, yet there are few effective software tools that automate team performance assessment and coaching.
The goal of this research is to develop an intelligent training framework where software agents are used to automate team performance assessment and coaching, with a focus on helping behavior. One of the design challenges in coaching for team training is that the performance of a team is affected by multiple factors, which include the quality of the team’s plan and each individual’s execution of the plan. To address this difficulty, our framework uses a two-phase training protocol that provides coaching for two phases: a mission planning phase and a mission execution phase. We adopt two user modeling approaches (overlay and error taxonomy) in Intelligent Tutoring System (ITS) for the two phases, according to different complexity involved in modeling expert behavior. In the planning phase, intelligent coaching feedback is generated based on an expert model for resource allocation. In the execution phase, coaching feedback is generated based on error taxonomy to assess team’s execution performance of their planned activities. Due to the broad scope of team training, this research stresses one important dimension of teamwork—the helping behavior among team members, in case of an unbalanced workload and resource distribution. Coaching feedback is provided in a debriefing session at the end of each mission execution with a goal of improving trainee performance during the next mission.
We have implemented the framework within a team-based agent architecture and applied it to train helping behaviors for a simulated command and control (C2) task. To evaluate the effectiveness of the agent-based team training approach, we designed and conducted a human subject experiment that applied the agent-based two-phase training protocol and provided teams in the experiment group with feedback generated by the coaching agents. Results have suggested that the coaching agents have a positive impact on trainees’ learning of how to effectively helping each other to achieve mission success in time-critical and complex task domains.
Teams allow workloads to be distributed among individuals with different expertise, and effective teamwork prompts improved team performance. There is a growing awareness that teamwork is vital to the success of many organizations, especially for complex tasks that impose high mental or physical demands that go beyond individual capacity—and that only effective communication and the collaboration of multiple members can ensure mission success.
With rapid technological advancement and intensive global competition, team training that has been one of the most important topics in psychology, has drawn more attention from researchers in various other disciplines to study team training technology that can be applied in either the military or the civilian sector.
The nature of teams is fundamental to understanding team performance and training. A team is defined as a distinguishable set of two or more people who interact dynamically, interdependently, and adaptively toward a common goal, who have each been assigned specific roles or functions to perform, and who have a limited life-span of membership  . The notion of a common goal in team definition distinguishes a team from other working groups who also have a certain degree of interaction and exchange of information or resources. Another unique characteristic of a team is that some kind of organizational structure needs to be imposed on the team members, and there must be some form of task dependency in order for them to coordinate and accomplish their shared objectives.
Training that aims to enhance team performance is more complex than the training of individuals to master certain task skills. Traditional one-on-one tutoring focuses on the modeling of the students at the individual level and provides feedback on how students can go through a problem-solving process better. While intelligent tutoring has become an effective method for a student to acquire individual skills, such as the Intelligent Tutoring System (ITS) application in learning math, algebra or programming language, its application in team training is still challenging. It is often observed that a team of experts does not make an expert team in many real-world scenarios.
More than 30 years of team training research has provided a great deal of information concerning the characteristics of a team, the dimensions of teamwork, and the factors that could influence team performance. Many studies have been designed to investigate factors that influence team performance outcomes, yet these studies have not provided a common answer to what exactly are the collective skills in the team performance process that training is supposed to improve. In this study we focus on arguably one of the most important dimensions of teamwork — helping behavior–– to guide our design and development of the intelligent training framework and its coaching assessment modules.
Team performance, defined by Nieva as goal-directed behaviors/function accomplished by a team in performing a task, has two major components:
individual-level task behavior and team-level task function  . It is an outcome of dynamic processes reflected in the coordination and communication patterns that a team develops over time. Individual expertise, as a portion of the solution to a complex problem, is necessary but not sufficient to achieve good team performance. Beyond the individual skills needed to perform their assigned subtasks, team members need to make decisions about how to communicate efficiently and to coordinate their work flow. One typical example of coordinated work flow is a team member proactively offering help to other members who are in need of support  .
Intelligent simulations provide interactive environments that are essential to automate team training. In such environments, trainees are able to interact with autonomous agents and other human team members and go through a set of realistic training scenarios where collective tasks need to be performed. Simulation-based training is less constraint by equipment cost or personnel availability that normally associated with human-to-human field training, and provides the potential for trainees to get training exercises almost anywhere anytime. To ensure trainees’ effective learning and transferring of the target knowledge and skills, intelligent agents are designed in such training environment to play the role of human coaches, monitoring trainees’ activities and providing feedback as necessary.
Intelligent tutoring system research provides a rich source of theories and practices that guide the development of automated tools in facilitating individual learning. However, it is not feasible to apply individual cognitive diagnosis, student modeling, and adaptive tutoring directly into team training due to the nature of team problem solving, such as the increased complexity, uncertainty, multi-layered knowledge spaces, and time stress. For example, initial attempts have been made to apply ITS concepts in developing intelligent team training technology, yet most of them focus on providing semi-automated instructions to reinforce human-based coaching, few succeeded in developing an intelligent training system where intelligent coach plays the role of a human coach/tutor  .
Planning has been long recognized as an effective way to ensure success in any endeavor that needs preparation. Planning before undertaking a task can result in efficiency enhancement and cost deduction, such as individual’s planning of a trip or budget or strategic planning in organization or tactical planning in the military. Reported in ITS literature, planning has been employed in guiding instructional design and discourse control [5, 6] , yet the benefit of student’s planning activities haven’t been recognized in the course of solving a problem. The lack of attention in ITS development on individual planning in ITS development might root in the nature of individual-based problem domains: 1) little indeterminacy is involved 2) the individual task can be highly procedural 3) the learning outcome highly depends on individual’s proficiency of a particular subject. Numerous such examples in ITS domain exist where planning doesn’t have significant impact on the learning outcome, such as solving a mathematical problem or trouble shooting a device. On the other hand, team problem solving always involves anticipation of task situation and adaptation to the possible future outcome. Thus for most team tasks, the demands of team-level decision making require a great focus on planning, which guides team’s proactive responding to potential interactions that might arise and plays a critical role in determining mission success.
Given the implications that planning has on training complex task skills, we proposed our “two-phase” training approaches that “divide and conquered” the team training problem at hand. In this approach, the training problem is decomposed into two sub-problems. Accordingly the whole training session is divided into a planning phase and an execution phase, each involves solving part of the training problem at different levels of complexity. During the planning phase, we allow trainees to practice on planning for their future mission and focus on their resource allocation skills in the context of helping overloaded team members. During the execution phase, trainees practice on execution of the detailed helping behavior, take advantage of the planned allocation strategies and try to adapt to changes evolved during the mission.
Besides the fact that the two-phase training has addressed the significance of planning activities as one important team training dimension, the distinction of the two sub-problems with different complexity also allow us to employ different ITS modeling techniques to suit the specific training needs in each phase. At planning phase, trainee’s planning task concentrates on knowledge at the abstract level, which is the allocation of team resources where that they are most needed. Yet in this phase, little detailed timing information is involved about when to initiate the specific action points to realize the planned resource allocation. Thus we could employ the overlay approach and extract the problem solution as an expert model that captures the desired planning strategies and provide planning phase feedback by comparing trainees’ actual planning strategies with the set of expert planning strategies. As more team dynamics are introduced in the execution phase, trainee actions might involve multiple layers of knowledge and skills. For example, the execution of their allocation plan entails team-level communication and coordination with the specific timing and outcome for each domain action. In the execution phase, however, it is not feasible to employ the overlay approach and build a comprehensive model of “correct” behaviors for an expert team. The characteristics of the execution-phase problem entail such difficulties: 1) Different than the traditional tutoring problem, there is no single right solution for the complex team task; 2) there are intensive interactions among team members and different combinations of team actions could lead to the same outcome. With the observation that the training needed at the execution level is not the provision of the entire solution set, we employ the error taxonomy approach to build trainee deficiency libraries at both individual level and the team level. To diagnose trainee deficiencies, we link multiple steps of event-based assessments to assess patterns of team’ helping behavior, which allows us to provide trainees with feedbacks that address their helping-related deficiencies.
The complexity of human team training has been intensively studied in cognitive science community. To design and develop the intelligent training framework, we also adopted a set of theoretically well-founded principles as reported from literature  . These principles are generic yet vital to guide the design of human team training for various domains.
Team training is hard problem to solve given the difficulties we described in the previous section (Section 1.2). Aiming to provide an automated training framework that addresses the team training problem, we need to carefully evaluate the scope of the research. Later in this section, we identify our target training goal by providing detailed characteristics of the training problem, including the nature of the team, the training environment, team task, target team dimension and the target skill level.
- The team being trained
We are looking at command and control (C2) teams that are responsible to accomplish complex team tasks by managing limit amount of resources within restricted time and space. Adequate amount of communication and collaboration are expected to facilitate team decision making and ensure success of the team mission. Examples of C2 teams include AWACS teams, air defense teams, first responder teams, incidence management teams, air traffic control teams, and NASA mission control center.
- The training environment
A realistic virtual environment is essential for simulation-based training and allows the modeling of individual operation, team collaboration and communication. Given a specific set of training goals, it is desirable that the simulation environment allows a range of adjustable parameters to configure the training scenarios. For example, in training of the helping behavior, it is necessary to have the situated collaboration context that requires team member’s helping each other. The team assessment and diagnosis capacities within the training framework need to be transferable to multiple training domains with similar team settings.
- Characteristic of task demands
As determined by the nature of C2 teams, the stress induced on team personnel may have played a significant role on the overall mission success while performing tasks in dynamic, time-critical, and uncertain environment  . Among incidents in warfare that were reported to exemplify the impact of stress factors for C2 teams, Vincennes tragedy has a tremendous implication on the urgent demand of training under stress, which refers to the tragedy when civilian Iran air flight 655 was mistakenly shot down by the US navy in 1988  . In this study, we introduced Dynamic Decision Making (DDD) where team members are overwhelmed with a large amount of incoming tasks during limited time period. Our training approach is based on the assumption that the introduction of stress in the tasks performed by trainees can reinforce the creation of valid simulation-based training environments that have positive impact on trainees’ dealing with stress in actual task-performance situations.
- Legitimacy of helping behavior
One of the main focus of this research is to study helping behavior, “a team skill that is at the heart of teamwork”  . In designing of our training scenario, we need to ensure a social setting where team member’s helping each other is legitimate and critical to achieve better team outcome. We focus on two aspects of a team that determine the need for helping— characteristics of the team’s task and the characteristics of the team composition. Team task can be quantified as the amount of workload assigned to each team member and team composition determines the resource distribution within team. The legitimacy of helping need is created when the workload assigned to a particular team member exceeds the amount of resources allocated to accomplish the task. As a specific case of the above described helping legitimacy, when team members share the same amount of resources (even resource distribution), the uneven distribution of task demand creates a clear and distinct need for team members’ helping each other  . Given that team’s incoming task load can be anticipated, observed or learned, the imbalance between the workload demand and resource allocation can be corrected. Helping behavior is highly involved in situations when team member need to recognize which teammate has been over loaded and provide direct assistance to compensate his/her lacking of resources.
- Skills to be acquired—Taskwork skills vs. teamwork skills
Taskwork skills refer to the ability of an individual to perform domain tasks that involves declarative and procedural knowledge at individual-level, such as individual’s awareness of a set of domain constraints specified by the task
(declarative) or a sequence of actions to perform a domain specific operation (procedural). Teamwork skills refer to the collective capacity of the team in terms of information sharing and coordinated decision making. Compared to taskwork skills that might be acquired individually, teamwork skills represent higher level cognitive tasks such as team situation awareness, active communication and well-coordinated joint acts and have to be trained within an interactive team context.
In this dissertation, we look at defining an agent-based team training framework and constructing coaching components to help trainee enhance performance in complex and time-stress domains where intense team interactions are required to ensure mission success. To design and implement the intelligent training system, we have two training phases each deal with a training sub-problem with different levels of complexity. In designing the assessment modules to diagnose trainee deficiencies in the second phase, we adopted the event-based training approach to link multiple steps of an event to assess patterns of helping behavior. Our design and implementation of the intelligent training tool allows further experimentation to validate training protocol and the effectiveness of coaching feedback.
The research addressed in this dissertation represents a rather limited yet in-depth coverage of many topics that concerns team performance and intelligent training. The two major team performance measures in training are problem diagnosis and skill development. Inspired by the psychological study about human team training, we diagnose the areas of difficulty in learning how to acquire team helping behavior. As an important dimension of teamwork, helping behaviors overlap with some other team dimensions, such as collaboration and coordination and involves a rich set of skill sets to be developed. To assist trainees with their skill development, we narrowed our research scope to be looking at C2 teams in a simulation environment. As from military teams in actual work situation, a collective skill set is identified to include command, control, communication, and coordination. Further, feedback is presented to a team at both the individual level and the teamwork level for optimal results.
We aim at building an intelligent coaching system that provides training feedback regarding a specific dimension of teamwork at both individual-level and team-level for a team of trainees. It helps them acquire appropriate knowledge and skills to perform highly interdependent tasks. We seek answers for two fundamental research questions about the design and evaluation of the intelligent team training system in this study: 1)
Regarding an important dimension of teamwork (helping behavior), how can we build an intelligent training system that enables software agents to play the role of a human coach and provide appropriate coaching feedback for a team of trainees to improve their performance? 2) What kind of training protocol and experimentation we design to evaluate the effectiveness of such training system for our current training objectives?
As we described earlier, we adopted the divided and conquer approach to decompose the training problem into a planning sub-problem and a coaching sub-problem in recognition of the significance of team planning and to be able to apply different ITS approaches according to different levels of problem complexity. Accordingly, during the course of training, a two-phase training protocol was employed as having a planning phase and an execution phase. The modeling of trainee on two different aspects of a complex problem and the design of the corresponding training protocol allows trainees to plan ahead for their mission before they carry out the mission. We had the assumption that trainees could achieve better mission performance by taking advantage of the anticipation of mission information during the planning phase and proactively adapt to changes during mission. During the after action review session, coaching feedbacks are presented for both trainee’s planning of resource allocation and their online execution of the team mission.
Mission Planning Mission Execution After Action Review
Pre-session In-session Post-session
Planning-phase Execution-phase coaching Review or feedback coaching agents agents within AIC framework provision
Through a set of
Overlay model Error taxonomy integrated graphical user
Assess trainee’s Assess individual and team Offline coaching feedback planning of level performance during regarding planning and
resource allocation trainee’s execution of mission execution
Generate Generate Planning Execution
planning-phase execution-phase coaching presentatiofeedback n presentatiofeedback n coaching feedback feedback
Figure 1 Two-phase Coaching within the Training Session Timeline
Figure 1 provides an overview of the training timeline of expected coaching and feedback sessions that occur in a complex tactical training environment. The whole training session consists of three sub-sessions, respectively pre-session, in-session and post-session. In the pre-session, trainees are allowed to plan about their mission-execution. Given a specific team plan, planning coach generates feedback regarding trainee’s planning of team resource allocation. The in-session includes the generation of execution-phase coaching feedback as well as the actual training simulation execution. During the post-session, coaching feedback generated by both planning phase and execution phase agents is presented to trainee with the goal to help them achieve better performance during the next training session.
For the mission planning phase, we built coaching agents that can assess trainees’ planning strategies and help them make better resource allocation plans for the next mission. Accordingly during the planning phase trainee, we provide trainees with information that helps them predict the upcoming mission, and allows them to make decisions about the allocation of resources during mission. The planning phase as part of the training problem is an abstraction of a trainee’s execution knowledge at the resource allocation level, with only rough timing information involved. The diagnosis capacity of the planning phase coaching agent is built on top of an overlay model where trainees’ planning strategies are viewed as a subset of the experts’ planning strategies. Trainee deficiencies are identified by comparing the student planning model with the expert planning model and feedback will be generated accordingly aiming at helping trainee make better plans during the next mission. Offline coaching feedback will be provided to trainee at the end of the whole mission execution regarding their resource allocation strategies compared to a set of expert resource allocation strategies.
For the mission execution phase, we built coaching agents that can assess trainees’ team performance during execution with a focus on their helping behavior and provide feedback regarding their performance deficiency both at the individual level and team level. The desired execution-level teamwork behavior involves intensive communication and interaction among team members and such a complex real-time problem does not imply a unique expert solution. To model trainee behavior during execution, it is not feasible to build a comprehensive overlay model to represent desired team behavior, thus a different type of assessment approach, error taxonomy model, is adopted. Within the error taxonomy model, a list of critical events regarding helping behavior is identified and a monitoring framework is developed for intelligent coaching agents to capture these critical events. We assess trainee’s helping pattern by linking multiple steps involved in each event category and reasoning about trainees’ deficiency regarding helping one another. During the after action review session, which is at the end of the whole training mission, trainees will also be given feedback regarding their helping behavior and other related individual performance deficiencies.
We also conducted human experiments to evaluate the effectiveness of our training protocol and the coaching feedback generated by the two-phase coaching agents within the intelligent training framework. The human experiment involved two groups of trainees that are both introduced to the two-phase training protocol, yet only the experiment group receives the coaching feedback generated by the two-phase coaching agents at the end of the training session. The underlying research hypothesis is that coaching agents within the agent-based intelligent training framework helped trainees achieve better team mission performance and specifically helped them learn how to proactively offer help to other members in a team context where such collaboration are desired and essential to accomplish the team tasks with high interdependency and complexity.
In this study, we focused on command and control teams that perform highly interdependent tasks in time-critical domain. Each member of the team has clear responsibilities, but an important strategy for enhancing team performance and processes is for each individual member to dynamically adapt to changes in the task environment, contributing their resources and coordination capabilities as necessary for the team to accomplish their mission successfully.
In designing the intelligent coaching agents within our training system, we built a performance assessment model that includes not only assessment of individual member’s task performance, but also a comprehensive team performance assessment that captures the collective variables that will significantly affect team performance during team’s mission execution. For the highly interdependent team tasks, our training goal is to help individual trainees learn how to improve inner-team collaboration capabilities to achieve teamwork effectiveness that couldn’t be achieved by team members individually.
We built our intelligent coaching agents on top of a multi-agent architecture. The generic coaching components can be applied to other teamwork-oriented domains where collaboration among members is key for overall mission success. Specifically we built a prototype training system that adds domain-specific coaching solutions for a command and control simulation. During the course of training, we applied the two-phase training protocols that allow trainees to first plan their mission execution and then carry out the plans they built as a team during mission. In planning phase, trainees will be given information that could be used to predicate characteristics of their incoming mission; in the execution phase, trainees carry out the plan they made as a team and make adjustments to adapt to the changes in the mission.
The intelligent training framework provides both planning phase coaching feedback and execution coaching feedback at end of each mission. The planning phase feedback addresses the issue of how to make a good plan that optimizes resource allocation of the overall team. Our training agents generate feedback about each individual team member’s planning deficiency comparing the actual team plan to a set of expert planning strategies. During a trainee’s execution of the mission, agents monitor human trainees’ actions, analyze data collected for critical collaborative events and provide feedbacks to trainees about their online deficiencies concerning individual task performance or their lack of the collaborative skills involved in team processes that are vital for the overall mission success.
Human experiments were conducted to validate our research hypothesis that in a simulation-based team training environment, intelligent coaching agents that provide offline feedback about team’s planning and execution of the mission have a positive influence on trainee’s achieving better collaborative process and performance outcome for tasks that requires coordinated interaction among team members.
In summary, this research has made contributions in the following aspects:
- Developed a generic team training framework that can be extended to other training domains for monitoring and giving feedback for a team of trainees to enhance both their team process and outcome.
- Developed intelligent coaching agents within the team training framework that implements an empirical training protocol for command and control teams in a military simulation domain.
- Conducted human subject experiments where teams of participants exercise with the above mentioned training protocol with or without the assist of intelligent coaching agents and validated our research hypothesis by comparing their performance process and outcome
INTELLIGENT COACHING AGENTS FOR ENHANCING HELPING BEHAVIOR IN HUMAN TEAMWORK