A Cognitive Model for the Generation and Explanation of Behaviour in Virtual Training Systems

A Cognitive Model for the Generation and Explanation of Behaviour in Virtual Training Systems
of 12
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  A Cognitive Model for the Generation andExplanation of Behaviour in Virtual TrainingSystems Maaike Harbers 1 , 2 , Karel van den Bosch 2 , Frank Dignum 1 , andJohn-Jules Meyer 1 1 Institute of Information and Computing Sciences, Utrecht University,P.O.Box 80.089, 3508 TB Utrecht, The Netherlands  maaike,dignum, 2 TNO Defence, Security & Safety, Kampweg 5, 3796 DE Soesterberg,The Netherlands  maaike.harbers, Abstract.  Instructors play a major role in many of the current vir-tual training systems. Consequently, either many instructors per traineeare needed, which is expensive, or single instructors perform highly de-manding tasks, which might lead to suboptimal training. To solve thisproblem, this paper proposes a cognitive model that not only generatesthe behaviour of virtual characters, but also produces explanations aboutit. Agents based on the model reason with BDI concepts, which enablesthe generation of explanations that involve the underlying reasons for anagent’s actions in ’human’ terms such as beliefs, intentions, and goals. 1 Introduction Virtual systems have become common instruments for training in organizationssuch as the army, navy and fire brigade. The following paragraph describes apossible training scenario of such systems. To practise his incident management skills, fire-fighter F receives training in a virtual environment. He and his virtual colleague G receive a call from the dis-patch centre informing that there is a fire in a house, and as far as known, there are no people in the house. While driving to the incident, they receive informa-tion about the exact location, scale of the incident, wind direction, etc. At the incident, F and G start unrolling and connecting the fire hoses, but suddenly,G sees a victim behind a window and runs towards the house. F did not see the victim and assumes that G is going to extinguish the fire. However, F sees that G is not carrying a fire hose, and wonders whether G forgot his hose, or ran tothe house for another reason. What should F do?  Virtual training systems provide an environment which represents those parts of the real world that are relevant to the execution of the trainee’s task. An interface  allows trainees to interact with the environment. In typical virtual training, atrainee has to accomplish a given mission, while a scenario defines the eventsthat occur. Despite the predefined scenario, the course of the training is notcompletely known, because the trainee’s behaviour is not exactly predictable.To provide effective training, it is important that the environment reacts tothe trainee’s actions in a realistic way, so that trainees can transfer the skillsobtained in the virtual to the real world.In most of the current systems for virtual fire-fighting training, virtual char-acters and other elements in the environment do not behave autonomously(e.g. [1]). Instead, instructors play a major role, managing changes in the en-vironment, e.g. the size of a fire, and impersonating other characters in thescenario, e.g. a trainee’s team-mates, police or bystanders, in such a way thatlearning is facilitated as much as possible. Besides controlling the environmentand behaviour of virtual characters, instructors provide instruction, guiding, andgive feedback to the trainee. As both of these tasks require ample attention of an instructor, more than one instructor is needed to train one trainee. Makinguse of even more instructors per trainee, however, is a costly and unpracticalsolution.A way to alleviate the tasks of an instructor is by using artificial intelligenceto perform (part of) the instructor’s tasks. A lot of research has been done onintelligent tutoring systems (ITS) (for overviews, see [2,3]). Successful applica-tions mostly concern the training of well-structured tasks, such as programmingand algebra [3]. Designing ITS for complex, dynamic, and open real-world taskstranspires to be difficult because it is not possible to represent the domain by asmall number of rules and the space of possible actions is large.Cognitive models are used to generate realistic behaviour of virtual charac-ters (e.g. [4]). For achieving simple behaviour, cognitive models are successfullyapplied, but generating complex character behaviour is more difficult. One of the problems is that in complex situations, it is not always clear why a characteracts the way it does. Even the system designers and the instructors, who shouldbe able to explain its behaviour, can only speculate about the character’s un-derlying reasons. Without knowing the motivation of a character’s actions, it ismore difficult for a trainee to understand the situation and learn from it.This paper presents an approach for virtual training with fewer instructors.We propose a cognitive model that does not only generate behaviour, but alsoexplains it afterwards, which should result in self-explaining agents. The paperhas the following outline: we first discuss what explanations of virtual charac-ters should look like (section 2), and then propose a cognitive model able togenerate such explanations (section 3). Section 4 discusses the two uses of themodel: generation and explanation of actions. Subsequently, requirements forthe implementation of the cognitive model (section 5) and related work (section6) are discussed. We conclude the paper with a discussion and suggestions forfuture research (section 7). We use the scenario at the beginning of the paper toillustrate the proposed principles and methods.  2 Self-explaining agents Early research on expert systems already recognized that the advice or diagnosesfrom decision support systems should be accompanied by an explanation [5–7].We transfer this idea to virtual training systems, where virtual characters shouldnot only perform believable behaviour, but also provide explanations about theunderlying reasons of their actions along with it. However, since there are manyways to explain a single event [8], the challenge is to develop a method to identifywhich explanations satisfactorily answer a trainee’s questions. The field of expertsystems distinguishes rule trace and justification explanations [5]. Rule traceexplanations show which rules or data a system uses to reach a conclusion.Justification explanations, in addition, provide the domain knowledge underlyingthese rules and decisions, explaining why the chosen actions are appropriate.Research shows that users of expert system often prefer justification to ruletrace explanations. Similarly, explanations of virtual characters should not onlymention how, but also why a certain conclusion has been reached.Humans use certain concepts when they explain their behaviour, they giveexplanations in terms of goals, beliefs, intentions, etc. Virtual agents should usesimilar vocabulary in order to produce useful explanations. To facilitate the gen-eration of explanations in ’human’ terms, agents can already make use of theseconcepts for the generation of their behaviour. Agents based on a BDI modelreason with concepts such as goals and beliefs. Moreover, it has been demon-strated that BDI agents are suited for developing virtual non-player charactersfor computer games [9]. Because BDI agents use relevant concepts for the gen-eration of their behaviour, explanations that give insight into their reasoningprocess are helpful to the trainee. For example, a BDI-based virtual charactercould explain its actions by referring to its underlying goals. 3 The cognitive model As argued in the previous section, we will use the BDI paradigm to developself-explaining agents. This section provides a model describing how beliefs, de-sires (goals), intentions (adopted goals) and actions relate to each other in ourapproach. The three basic elements in the model are goals, actions and beliefs.As illustrated in figure 1, a goal can be divided into sub-goals or achieved byexecuting a single action. Beliefs relate to the connections between goals and sub-goals, and goals and actions. The three subsections each discuss one of them,and explain the relations to the other elements in the model more extensively. 3.1 Goals in the model A goal describes (part of) a world state. All possible goals of an agent are presentin its cognitive model and depending on the current world state, one or more of its goals become active. An agent’s knowledge about the current world state isrepresented in its beliefs, determining which goals become active. For example,  Fig.1.  The relation between goals, actions and beliefs a fire-fighter would only adopt the goal to extinguish a fire, if he believed thatthere (possibly) was one.Goals relate to their sub-goals in four different ways. The  first  possibilityis that all sub-goals have to be achieved in order to achieve the main goal,and the order of achievement is not fixed. For instance, the goal to deal with afire in a house is divided into the sub-goals to extinguish the fire, and to savethe residents. The completion of both sub-goals is essential to the achievementof the main goal, and both orders of achievement are possible. In the  second relation, the achievement of exactly one sub-goal leads to the achievement of the main goal. The different sub-goals exclude each other in the sense that if anagent adopts one of the sub-goals, the other sub-goal(s) cannot simultaneouslybe active goals. For example, a main goal to extinguish a fire has the sub-goalsto extinguish the fire with water and with foam. A fire-fighter has to choosebetween water and foam, but he cannot use both. In contrast to the first example,neither of the sub-goals is necessary, but both are possible ways to achieve themain goal. In the  third  goal/sub-goal relation, the achievement of one sub-goalleads to the achievement of the main goal, but the different sub-goals do notexclude each other. For example, sub-goals of the goal to find a victim are toask other people, and to search the victim yourself. Different from the choicebetween water and foam, in this example an agent can decide to adopt bothstrategies. In the  fourth  relation, for achieving the main goal, all sub-goals haveto be achieved in a specific order. For instance, the goal to rescue a victim in aburning house has the sub-goals to find the victim and to bring him to a safeplace. The main goal is only achieved if the two sub-goals are achieved in thecorrect order.The four goal/sub-goal relations are indicated by adding subscripts to thesub-goals, denoting the relation to their main goal. The different sub-goals are G and , G xor , G or  and G seq − i , in the order as discussed in the previous paragraph,where  and  ,  xor   and  or   refer to the logical operators, and  seq   refers to sequential.The  i   in G seq − i  stands for the position of a sub-goal in the sequence of sub-goals.It should be noted that the logical operators are biased, as some of the sub-goalsare adopted more often than others. Each of the sub-goals can also be a maingoal with new branches of sub-goals starting from it. Thus, the name of a goalis always derived from the relation to its main goal and the other sub-goals of   that main goal. Some goals can have more than one sub-script, i.e. if they relateto more than one main goal in different ways. 3.2 Actions in the model Goals are specified by sub-goals, which are specified by new sub-goals, etc. Atsome point in the goal tree, goals are no further specified and can be achievedby performing one single action. Our model does not allow connecting morethan one action to a goal for simplicity reasons. The different relations betweentwo levels can be expressed at the goal/sub-goal level, so distinguishing differentgoal-action relations is unnecessary.Basic actions should be carefully chosen in the model. If two actions arealways followed by each other, e.g. the action to connect a fire hose and to openthe tap, these two actions can be represented as one action in the model. Adivision of two actions (and thus two sub-goals because goals relate to only oneaction), while they are never used separately, is unnecessary. The determinationof the smallest distinguishable actions does not only depend on the execution of actions, but also on the explanation of actions. 3.3 Beliefs in the model In our model, an agent’s beliefs determine which of its goals become active, andthus which actions it takes. Agents have two types of beliefs: general and ob-servation. General beliefs involve all them of which the truth does not dependon a particular situation, for example, the belief that combustion requires oxy-gen. Observation beliefs are context-specific and involve them about an agent’scurrent situation. Observation beliefs are only true in a particular situation, forexample, the belief that there is somebody in the burning house. It should benoted that observation beliefs are often interpretations of observations, whichmeans that additional knowledge has been added. For example, the belief that ahouse is burning is an interpretation of observations such as smoke, heat, soundsof fire, etc. In general, it is hard to draw a line between observations and interpre-tations, but we will take a practical approach and consider such interpretationsas observation beliefs.In simple worlds, e.g. a block world, it is possible that an agent has aninternal representation of all possible information about the current state of theenvironment. In more complex worlds, such as in virtual training systems, theenvironment contains lots of rapidly changing information and representing allof it is unfeasible. In our model, only beliefs that influence the agent’s actionsor that are needed for explaining them are made explicit. 4 Uses of the model In this section, we illustrate how our model can be used for the generation andthe explanation of behaviour. Therefore, we use the scenario at the beginning of 
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks