Virtual training systems are increasingly used for the training of complex, dynamic tasks. To give trainees the opportunity to train autonomously, intelligent agents are used to generate the behavior of the virtual players in the training scenario. For e ective training however, trainees should be supported in the re ection phase of the training as well. Therefore, we propose to use self-explaining agents, which are able to generate and explain their own behavior. The explanations aim to give a trainee insight into other players' perspectives, such as their perception of the world and the motivations for their actions, and thus facilitate learning. Our project investigates the possibilities of self-explaining agents in virtual training systems, and the e ects on learning.
Virtual training is used to train people for complex, dynamic tasks in which fast decision making is required, e.g. crisis management, military missions or reghting. In typical virtual training, a trainee has to accomplish a given mission and therefore he has to interact with other virtual players, e.g. team-members, opponents, or neutral participants. Currently, in most virtual training these are controlled by other trainees or instructors. However, using intelligent agents instead of humans gives trainees more exibility to train where and whenever they want, and it reduces costs. Fire- ghters could for example train during a night shift, when they spend most of their time waiting for an alarm.
Intelligent agents can only (partly) replace humans if they are able to
generate believable behavior, which might be complex. Moreover, trainees should be
supported to re ect on the training because that promotes learning [
In this paper, we present a PhD project issuing the topic of self-explaining agents in virtual training. In section 2, we discuss some related work, and in section 3 we give a formulation of our research question. Then, we provide a more detailed discussion on our approach and the results achieved so far in section 4. We end the paper with a conclusion and an outline of future research in section 5.
A lot of research has been done on intelligent tutoring systems (ITS) [
A few proposals for self-explaining agents in virtual training systems have
been made. The rst called Debrief [
A more recently developed account of self-explaining agents is the XAI
explanation component [
A second version of the XAI system [
The previous section showed that ITSs usually support the learning of wellstructured tasks, in which it is clear what actions are right and wrong. In the virtual training systems we focus on however, training tasks can be achieved in many di erent ways and require another type of feedback than provided by ITSs. Self-explaining agents might be a good alternative, but we believe that the existing accounts lack some crucial aspects. They either just give explanations about the agent's physical state, or they derive information about the agent's mental state from its behavior or the simulation. We believe that an agent's behavior can best be explained by its actual underlying motivations, i.e. information about its mental state, and that the explanation component thus should have direct access to this information and not on their e ects. To solve the shortcomings in the current solutions, the PhD project presented in this paper addresses the following research question: How can we develop self-explaining agents and how can they be applied in virtual systems to support training? The question is two-fold, the rst part is a technical question, and the second part focuses more on educational aspects. In the remainder of this paper we discuss our methodology, and the results achieved so far. 4
To obtain direct access to an agent's reasons for executing an action, we believe
that behavior explanation should be connected to the generation of behavior.
The deliberation steps that are taken to generate an action can also be best
used to explain that action, and when these deliberation steps are
understandable, the explanations should be as well. To obtain understandable deliberation
steps, the agent's reasoning elements should have some level of abstraction. For
instance, the description "an agent is opening a door" is more useful for
understanding its behavior than "an agent is moving object x from position (x1,y1,z1)
to (x2,y2,z2)". We have chosen to use a BDI-based (beliefs desires intentions)
approach [
We have chosen for the BDI approach because it matches the way humans
give explanations. Humans adopt a certain 'stance' or 'mode of construal' for
explaining and predicting phenomena, and di erent stances are chosen to explain
di erent phenomena [
The BDI approach de nes an agent's reasoning elements, but it does not tell
how actions can be generated from an agent's goals and beliefs, i.e. how planning
works. For an account of planning, we have looked at the GPGP (generalized
partial global planning) approach [
Conform the GPGP approach, we structure the possible goals, plan and actions of an agent in a task hierarchy. The task at the top of the hierarchy is an agent's goal, the subtasks possible plans for reaching that goal, and the leaves are the agent's actions. Consequently, for each of the agent's goals, a task hierarchy is de ned. The actions that an agent executes to achieve a goal depend on three aspects (explained in the next paragraph): its beliefs, the nature of the task-subtask relation, and its preferences.
First, the GPGP model is designed for a team of agents, but we take a single agent perspective. Therefore, in our model the beliefs of a single agent can be added to the task hierarchy, to form the conditions that determine which tasks can possibly be executed. Second, three task-subtask relations can be distinguished. A task can be executed when: { All subtasks are executed { One subtask is executed { All subtasks are executed in a speci c Order Third, the agent's preferences determine the order in which subtasks are executed in an All-relation, and which subtask is executed in an One-condition.
Figure 1 shows a the model of a simple re- ghting agent. The agents main goal is to handle the incident, and it has several plans available to achieve this goal. Its current beliefs determine how the agent 'walks through' the hierarchy. The rst step is to choose between saving a victim (if the agent beliefs that there is an actual victim), or extinguishing a re (if the agent beliefs that there is a re and no victim because saving victims is preferred over extinguishing res). For saving a victim, it rst has to search the victim and then carry it away. For extinguishing a re it can either use water or foam, dependent on its beliefs about the availability of water and foam.
The same agent model can also be used for the generation of explanations about the agent's actions. The actions that are the result of an agent's deliberation process can be explained by the beliefs, goals and reasoning steps that were in involved in the process. For instance, extinguishing a re with foam could be explained as follows.
I wanted to handle the incident,
and I believed there was a fire and no victim,
therefore I wanted to extinguish the fire,
and I believed there was foam and no water,
therefore I used foam
Such explanations can become quite long, which is not desired [
For the implementation of our agent model it is required that the agent's
reasoning elements and its deliberation steps can be explicitly represented in
the programming language. Second, an agent needs to have access to this
information. We have chosen to implement our model in the agent programming
language 2APL [
We argued that virtual training of complex and dynamic tasks requires intelligent agents that can provide explanations for their behavior in such a fashion that it helps trainees to improve their understanding. So far, we have developed an agent model capable of the generation and explanation of behavior, and we have made an implementation of the model.
We are currently reviewing literature on cognitive behavior research to determine which information people nd most useful in explanations. Based on the outcome of this study we want to develop lters which select the most useful information out of longer explanations. Furthermore, we want to extend the agent model with factors like emotions, personality or social contracts, which may also in uence an agent's behavior. Explanations referring to these properties may help trainees to become more sensitive and understanding to them. The next step will be to connect the self-explaining agent to an existing virtual training system, and perform user experiments.
We believe that our approach can create a learning tool that is currently not existing, and which ful lls the requirements of autonomous training of complex and dynamic training tasks. Our goal in this project is to demonstrate that the self-explaining agents we are developing deliver appropriate and useful explanations, leading to improved learning.
Acknowledgements This research has been supported by the GATE project, funded by the Netherlands Organization for Scienti c Research (NWO) and the Netherlands ICT Research and Innovation Authority (ICT Regie).