assistant agents to advice users in hybrid structured 3d virtual environments

10
COMPUTER ANIMATION AND VIRTUAL WORLDS Comp. Anim. Virtual Worlds 2014; 25:495–504 Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/cav.1598 SPECIAL ISSUE PAPER Assistant agents to advice users in hybrid structured 3D virtual environments Pablo Almajano 1 *, Maite Lopez-Sanchez 2 , Inmaculada Rodriguez 2 and Tomas Trescak 3 1 IIIA - CSIC, University of Barcelona, Barcelona, Spain 2 University of Barcelona, Barcelona, Spain 3 University of Western Sydney, Sydney, Australia ABSTRACT Hybrid structured 3D Virtual Environments model serious activities in immersive 3D spaces, where participants are human and software agents, and their interactions are regulated by an Organization Centered Multi-Agent System (OCMAS). In this context, both OCMAS social model and the tasks that users need to accomplish can be rather complex, and thus, users may benefit from having an assistance service. Hence, we propose personal assistant agents (PA), which, based on knowledge about the OCMAS specification and current system state, provide the user with an advice (a plan) to achieve her or his goal. Additionally, we implement this service with PLAN- EA, an Extension of the A algorithm that generates plans for a user whose actions may depend on other users’ actions. Thus, PAs provide plans that do not only include assisted user actions but also other users’ ones. We illustrate our approach by means of v-mWater—an online water market—and make a comparative analysis, with and without assistance, where efficiency—in terms of number of user actions—shows an improvement (7 vs 10.8), efficacy—percentage of completed tasks—also improves (93% vs 77%), and assistance’s overall satisfaction is positive. Copyright © 2014 John Wiley & Sons, Ltd. KEYWORDS personal assistant agents; social model; structured 3D virtual environment; OCMAS planning *Correspondence Pablo Almajano, IIIA - CSIC, University of Barcelona, Barcelona, Spain. E-mail: [email protected] 1. INTRODUCTION The blending of digital technologies, such as artificial intel- ligence, interactive systems, 3D interfaces and the Internet, is transforming the way people interact with computers. This blending is enabling new services for users, favoured by huge advances in the development of graphics and networking. In particular, multi-user online 3D virtual environments (VE) provide users with a collaborative space not only for entertainment and socialisation but also for developing ‘serious’ activities like learning, negotiating and shopping. Most often, these serious activities require the execu- tion of complex tasks, involving different subtasks, and highly regulated interactions. Nevertheless, standard 3D VE lack of mechanisms to structure (regulate) users’ inter- actions. An Organization Centered Multi-Agent System (OCMAS) [1] constitutes a powerful tool that can serve this purpose. Specifically, the marriage of 3D VE and OCMAS allows the participation of both human and software (SW) agents in a so-called hybrid structured system. However, its intrinsic complexity constitutes a serious obstacle for the participant system comprehension, which is seldom aware of all its runtime property values. In this context of hybrid structured VEs, we propose Per- sonal Assistant agents (PA), which, based on knowledge about the OCMAS specification and current state, support participants to understand the social model underlying such a specification and to perform complex tasks in a seamless way. We describe the overall system as a Hybrid Assisted Structured 3D Virtual Environment. Personal assistants provide different general assistance services. In a previous work, we presented an Information Service for SW agents [2]. This paper focuses on the for- malisation, operationalisation, and evaluation of the Advice Service for human users. Through this advice service, a PA provides the user with a plan (i.e. a set of actions compatible with the OCMAS specification) to accomplish her or his goal (task). PA embodiment is an ‘Angel’, which may accompany the user during the interactive experience and is available under request. Copyright © 2014 John Wiley & Sons, Ltd. 495

Upload: tomas

Post on 01-Apr-2017

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Assistant agents to advice users in hybrid structured 3D virtual environments

COMPUTER ANIMATION AND VIRTUAL WORLDSComp. Anim. Virtual Worlds 2014; 25:495–504

Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/cav.1598

SPECIAL ISSUE PAPER

Assistant agents to advice users in hybrid structured3D virtual environmentsPablo Almajano1*, Maite Lopez-Sanchez2, Inmaculada Rodriguez2 and Tomas Trescak3

1 IIIA - CSIC, University of Barcelona, Barcelona, Spain2 University of Barcelona, Barcelona, Spain3 University of Western Sydney, Sydney, Australia

ABSTRACT

Hybrid structured 3D Virtual Environments model serious activities in immersive 3D spaces, where participants are humanand software agents, and their interactions are regulated by an Organization Centered Multi-Agent System (OCMAS). Inthis context, both OCMAS social model and the tasks that users need to accomplish can be rather complex, and thus,users may benefit from having an assistance service. Hence, we propose personal assistant agents (PA), which, based onknowledge about the OCMAS specification and current system state, provide the user with an advice (a plan) to achieve heror his goal. Additionally, we implement this service with PLAN-EA, an Extension of the A� algorithm that generates plansfor a user whose actions may depend on other users’ actions. Thus, PAs provide plans that do not only include assisted useractions but also other users’ ones. We illustrate our approach by means of v-mWater—an online water market—and makea comparative analysis, with and without assistance, where efficiency—in terms of number of user actions—shows animprovement (7 vs 10.8), efficacy—percentage of completed tasks—also improves (93% vs 77%), and assistance’s overallsatisfaction is positive. Copyright © 2014 John Wiley & Sons, Ltd.

KEYWORDS

personal assistant agents; social model; structured 3D virtual environment; OCMAS planning

*Correspondence

Pablo Almajano, IIIA - CSIC, University of Barcelona, Barcelona, Spain.E-mail: [email protected]

1. INTRODUCTION

The blending of digital technologies, such as artificial intel-ligence, interactive systems, 3D interfaces and the Internet,is transforming the way people interact with computers.This blending is enabling new services for users, favouredby huge advances in the development of graphics andnetworking.

In particular, multi-user online 3D virtual environments(VE) provide users with a collaborative space not onlyfor entertainment and socialisation but also for developing‘serious’ activities like learning, negotiating and shopping.

Most often, these serious activities require the execu-tion of complex tasks, involving different subtasks, andhighly regulated interactions. Nevertheless, standard 3DVE lack of mechanisms to structure (regulate) users’ inter-actions. An Organization Centered Multi-Agent System(OCMAS) [1] constitutes a powerful tool that can serve thispurpose. Specifically, the marriage of 3D VE and OCMASallows the participation of both human and software (SW)agents in a so-called hybrid structured system. However, its

intrinsic complexity constitutes a serious obstacle for theparticipant system comprehension, which is seldom awareof all its runtime property values.

In this context of hybrid structured VEs, we propose Per-sonal Assistant agents (PA), which, based on knowledgeabout the OCMAS specification and current state, supportparticipants to understand the social model underlying sucha specification and to perform complex tasks in a seamlessway. We describe the overall system as a Hybrid AssistedStructured 3D Virtual Environment.

Personal assistants provide different general assistanceservices. In a previous work, we presented an InformationService for SW agents [2]. This paper focuses on the for-malisation, operationalisation, and evaluation of the AdviceService for human users. Through this advice service, aPA provides the user with a plan (i.e. a set of actionscompatible with the OCMAS specification) to accomplishher or his goal (task). PA embodiment is an ‘Angel’, whichmay accompany the user during the interactive experienceand is available under request.

Copyright © 2014 John Wiley & Sons, Ltd. 495

Page 2: Assistant agents to advice users in hybrid structured 3D virtual environments

Assistant agents to advice users in hybrid structured 3D VE P. Almajano et al.

We illustrate and evaluate our approach by means ofv-mWater, a regulated virtual environment for the tradingof water rights. Our evaluation results assess that, when-ever requested, this Advice Service provides comfortableand clear guidance that facilitates task accomplishment. Infact, its usage reduces both the percentage of task failuresand the number of user performed actions, when comparedto a previous system, we evaluated without assistance [3].Assuming we think in the same type of actions (i.e. theones defined in the system specification), we consider thenumber of actions an efficiency measure: the less the num-ber of actions to do the task, the more directly users reachthe goal. The smaller number of actions also implies lesscognitive load on the user, and a reduction of the taskperceived complexity.

2. RELATED WORK

A number of works in the literature focus on (intelli-gent) virtual agents providing users with assistance. In theline of recommender systems, Dong, et al. [4] propose aReviewer’s Assistant to give a product reviewer optionalsuggestions. This assistant is a web browser plug-inthat employs different data mining strategies. Anotherrecommender system [5] uses a multi-agent system (MAS)populated by PA agents. A rule driven PA provides the userwith talks announcements. To do so, it exploits semanticweb data and the user profile. In these works, assis-tants serve one (assisted) single-user not involved in anyorganisation. As a remarkable difference, our PAs operatein a multi-user and structured social scenario, characterisedby complex interleaved interactions.

Lujak, et al. [6] develop Emergency Medical Assistance(EMA), a medical emergencies scenario modelled as anOCMAS, which employs web services and a 2D UserInterface (UI) to assist participants (e.g. patients, ambu-lances and medical professionals) in their coordination.RADAR [7] is another MAS with a Multi-Task Coordina-tion Assistant that observes experts and learns models toassist office workers cope with email overload. Its adviceincludes task ordering suggestions and warnings when theuser’s behaviour differs significantly from the expert’s one.They propose a 2D task-management UI. Electric Elves [8]also developed specific software personal assistants (SPA),but they were for project activities coordination and exter-nal meetings organisation in a real environment. Unlike

our shared and collaborative 3D space with embodied PAs,these works feature 2D user interfaces where the PAs arenon-embodied.

A recent work has presented a coalition planning assis-tant agent in a peacekeeping problem domain [9]. In thisscenario, the user moves in an n � n grid towards a prede-fined subset of goal cells, and norms represent prohibitionsand obligations on visiting a cell. The assistant agent is ableto discover the particular user’s goal and execute actions(e.g. send escort request) to avoid user norm violations. Asa difference, our PA provides users with a plan (actions) toachieve an indicated goal. Moreover, their agents executeindependent actions to ensure norm compliance of usermovements. Finally, assistance has been used to supportusers’ physical navigation in 3D virtual environments, suchas the Virtual Theatre [10], where a non-embodied agentuses environment knowledge to indicate the user places tovisit; and the Virtual Museum [11], where a 3D characterpresents the user a precomputed guided tour to follow. OurPA is an interactive 3D character that is always available tothe user and helps her or him to achieve a given task.

3. PERSONAL ASSISTANT

This section first introduces the knowledge that a PA agentneeds to provide an advice (plan) to the human user. Next,it describes the operationalisation of the Advice Service.It illustrates them with v-mWater, an example scenarioin the agriculture domain that models an electronic mar-ket for trading water rights (i.e. the right to use water forirrigation).

3.1. Agent Knowledge: Social Model

Our structured 3D VE integrates a 3D virtual world(VW) and an OCMAS. On the one hand, the 3D inter-face facilitates the interaction of human users. On theother hand, the OCMAS infrastructure (an ElectronicInstitution [12]) regulates real-time participants’ interac-tions, where participants can be both human and softwareagents. A PA creates a plan for the user employing, asknowledge, the static specification of the OCMAS socialmodel (Spec) and the system current (dynamic) state (Sc).Particularly, it considers the social structure and social

Figure 1. Extract of the static specification of v-mWater social model.

496 Comp. Anim. Virtual Worlds 2014; 25:495–504 © 2014 John Wiley & Sons, Ltd.DOI: 10.1002/cav

Page 3: Assistant agents to advice users in hybrid structured 3D virtual environments

P. Almajano et al. Assistant agents to advice users in hybrid structured 3D VE

conventions (SocStr, SocConv 2 Spec) and their runtimevalues (RtSocStr, RtSocConv 2 Sc).

Social structure (SocStr) defines the properties associ-ated to user agents (AgP), the roles (Rol) that participantsenact, and their relationships (Rel). Figure 1 depicts anextract of v-mWater specification concerning seller par-ticipants. Its SocStr (at the top) specifies two partici-pant properties—their location (loc) and goal—and twoparticipant roles—seller (s) and market facilitator (mf ).

SocConv D hActiv, ActivRel, Proti (1)

ActivRel D hTra, Movi (2)

mov D hmRol, ori, des, mTi (3)

prot D hProtP, ProtSpeci (4)

ProtSpec D hProtC, Nod, Illoi (5)

ProtC D fhrolC, min, maxig (6)

nod D hEnterRol, ExitRoli (7)

illo D hsRol, rRol, cM, ori, desi (8)

Social conventions (Equation (1)) stand for the ‘rules ofthe game’. Participants (roles) gather in activities (Activ).There, they can perform tasks by following protocols(Prot) or move to other related activities (ActivRel). In ourexample of Figure 1, seller participants can enter/leave thesystem in the Initial/Final activity, ask for market informa-tion in the Waiting&Info activity and register a water rightin the Registration. Activities relationships (ActivRel) inEquation (2) constraint the flow of roles among activities.Transitions (Tra) are intermediate locations meant for usersynchronisation. Movements (mov 2 Mov, Equation (3))are participant actions that change their location. They aredefined by (i) its role (mRol); (ii) the transition and activityit connects (ori, des 2 fTra [ Activg, ori being origin anddes destination); and (iii) its type (mT 2 {“new”, “enter”,“exit”}, meaning create and enter a new activity, enter andexit an existing activity). Figure 1 represents a directedgraph where nodes correspond to both activities (rect-angles) and transitions (diamonds), and edges (labelledarrows) represent allowed movements between them.

Protocols structure user interactions within activities.Equation (4) defines a protocol (prot 2 Prot) as a setof properties (ProtP) and its specification (ProtSpec).ProtSpec (Equation (5)) is a finite state machine, wherenodes (Nod) correspond to the states and illocutions (Illo)to state transitions. Figure 1 depicts protocols’ nodes as cir-cles and illocutions as labelled arrows. Those nodes wherethe protocol execution starts and ends are named initial andfinal, respectively. Moreover, the protocol capacity (ProtCin Equation (6)) defines a set of tuples that bound the num-ber of participants that enact each role (rolC): min sets theminimum number required to open the protocol, whereasmax sets the maximum number of allowed participants in

this protocol. We can also specify, for each protocol node(nod 2 Nod) in Equation (7), the list of roles that can join(EnterRol) and leave (ExitRol) the activity when it is at thisprotocol state. For example, in the Registration activity ofFigure 1, users playing seller role are allowed to enter (Cs)and exit (�s) whenever its protocol state is node n1.

Finally, illocutions are messages that participants inter-change. They act as transitions of the finite state machine,so they imply a change in the protocol execution state(node). Thus, for instance, any illocution uttered at aninitial node opens the protocol. An illocution (illo inEquation (8)) formalises a participant action by specifyingthe following: (i) the role of the sender (sRol) or all, incase anybody can send it; (ii) the role of the receiver (rRol)or all, if it is public and everybody will receive it; (iii)the message content (cM), conforming an ontology; (iv)the origin node (ori) where the message can be sent; and(v) the destination node (des), the node reached once theillocution is successfully uttered. In Figure 1, illocutionregister (s, mf , hw, pi) in Registration specifies the requestfrom a seller (s) to a market facilitator (mf ) the registrationof a water right (w) to a given price (p). It can be utteredwhen the protocol’s state is node n1 and, once performed,the protocol state will transit to n2.

Execution State (Sc) contains current runtime values ofdynamic organisational elements, such as actual partici-pants (RtAg 2 RtSocStr) and running activities (RtActiv 2RtSocConv).

3.2. Advice Service: Operationalisation

Activities in the social model specify user interactions sothat certain tasks can be performed (e.g. users can registerwater rights in the Registration activity). Whenever a useraims to perform one of such tasks, she or he can request anadvice service to her or his PA. This PA will then respondwith a plan (pl D fa1, : : : , amg) consisting of a sequenceof m actions (movements and illocutions). The providedadvice (plan) will conform to the social model specificationand, if executed at the current system state, will lead theuser to her or his goal of performing the task.

Equation (9) formalises this advice service. When auser request (Req) is received, the PA uses social modelspecification (Spec) and current state (Sc) to provide aresponse (Res). Both Req and Res are messages, whichcontain a sender, a receiver and a content. In the former(Equation (10)), the user asks for PA’s advice to perform atask. In the latter (Equation (11)), the PA answers the userwith a plan (pl).

Advice : Req � Spec � Sc ! Res (9)

Req D hrtAg, PA, taski (10)

Res D hPA, rtAg, pli (11)

A PA is graphically represented as an Angel bot, aninteractive non-player 3D character. Figure 2 shows a

Comp. Anim. Virtual Worlds 2014; 25:495–504 © 2014 John Wiley & Sons, Ltd. 497DOI: 10.1002/cav

Page 4: Assistant agents to advice users in hybrid structured 3D virtual environments

Assistant agents to advice users in hybrid structured 3D VE P. Almajano et al.

Figure 2. Mary’s avatar with her Personal Assistant in the 3D VW.

snapshot of our 3D UI, with user Mary (female user avatar)and her PA (Angel bot). PA always accompanies the userduring the interactive experience so it is always availablefor her. Nevertheless, she can activate (show) her PA, inter-act with it whenever she needs help or deactivate (hide) itotherwise.

User-PA interaction has been implemented in the fol-lowing way: (i) A user performs the action touch to her orhis PA in the 3D VW to start a help request. Then, the PAshows an option dialogue (on the right of Figure 2) wherethe user can select one of the available Advice Services.(ii) The PA sends a plan to the user as a note card (onthe left of Figure 2), which is a document where the plan,initially generated using a notation only readable by SWagents, is written in natural English language to facilitateuser comprehension.

Based on the social model specification, the designeridentifies those complex tasks that will be automaticallyassisted. Particularly, in v-mWater PAs help (i) sellers toask for information about market prices and to registerwater rights and (ii) buyers to participate in auctions. Nextsection details the planning process, which is general forany social model specification.

4. ORGANIZATION CENTEREDMULTI-AGENT SYSTEM PLANNING

As previously stated, PAs compute plans based on both thestatic OCMAS social model specification (Spec) and cur-rent (dynamic) system state (Sc). This knowledge allowsto explore the search space by expanding a directed plan-ning tree to compute the path towards a goal state (i.e. theone reached once a user accomplishes a task). Nodes ofthe planning tree represent different system states, whereasedges correspond to possible participant actions. The goalis expressed in terms of desired values for state runtimeproperties (e.g. the location of the user). Notice though,that, because we consider a multi-user scenario with actiondependencies, we need to reckon with different plansexecuted by different participants.

Specifically, a PA computes a plan for a participantby invoking the recursive function PLAN-EA.Spec, rtAg,PlSc/ with the following parameters: (i) its corresponding

runtime agent (rtAg), including its goal (rtAg.goal); (ii) thesocial model specification (Spec); and (iii) current plan-ning state (PlSc), a working copy of Sc. Considering ourmulti-user scenario, we propose to implement PLAN-EA asan Extension of A� [13] (with an admissible heuristic) thathandles action dependencies by providing plans that areextended with other users’ plans. By action dependencies,we mean actions whose success depend on other users’sactions. For example, Mary needs an activity to be openbefore she can move into it, or she cannot utter an illocutionif there are no receivers. Thus, plans for opening activitiesor including receivers are associated to her plan, so she willbe aware that she has to wait for other participants to exe-cute these plans before she can successfully perform herplanned action.

On the one hand, to utter an illocution in an activityrtActiv requires a sender, a receiver and, if applicable,the min number of participants from the protocol capacity(Equation (6)). For each missing participant at rtActiv, wecall ASENT.Spec, rol, rtActiv, PlSc/ in Algorithm 1, whichreturns an associated plan for entering the activity. It usesfunction AGROLNACT to select participants located out-side rtActiv and enacting role rol. Then, for each selectedrtAg, it computes a plan to join rtActiv by setting its goalto be this location. ASENT returns the plan having theminimum cost.

On the other hand, to enter an activity, activ requires theactivity to be (1) created and (2) opened to the participantagent. Regarding (1), ASCREATE.Spec, activ, PlSc/ inAlgorithm 2 computes the associated plan for creatingactiv. It first invokes CANCREATE to select agents (RtAg)

Algorithm 1 ASENT.Spec, rol, rtActiv, PlSc/

TmpPls ;RtAg AGROLNACT(Spec, rol, rtActiv, PlSc)for all rtAg 2 RtAg do

rtAg.goal frtAg.loc DD rtActivgpl PLAN-EA(Spec, rtAg, PlSc)TmpPls TmpPls [ fplg

end forreturn MINCOST(TmpPls)

498 Comp. Anim. Virtual Worlds 2014; 25:495–504 © 2014 John Wiley & Sons, Ltd.DOI: 10.1002/cav

Page 5: Assistant agents to advice users in hybrid structured 3D virtual environments

P. Almajano et al. Assistant agents to advice users in hybrid structured 3D VE

Algorithm 2 ASCREATE.Spec, activ, PlSc/

TmpPls ;RtAg CANCREATE(Spec, activ, PlSc)for all rtAg 2 RtAg do

rtAg.goal frtAg.loc DD new.activ/gpl PLAN-EA(Spec, rtAg, PlSc)TmpPls TmpPls [ fplg

end forreturn MINCOST(TmpPls)

Algorithm 3 ASOPEN.Spec, rol, rtActiv, PlSc/

TmpPls ;Nod ENTERNODS(Spec, rtActiv, rol)RtAg AGACT(rtActiv, PlSc)for all nod 2 Nod do

for all rtAg 2 RtAg dortAg.goal frtActiv.state DD nodgpl PLAN-EA(Spec, rtAg, PlSc)TmpPls TmpPls [ fplg

end forend forreturn MINCOST(TmpPls)

currently enacting a creator role for activ. Second, for eachrtAg, it sets the creation goal and calls PLAN-EA to haveits plan. Third, it returns the minimum cost plan. As for(2), ASOPEN.Spec, rol, rtActiv, PlSc/ in Algorithm 3 calls:ENTERNODS to obtain all nodes (Nod), where role rol canenter; AGACT to obtain all activity participants (RtAg).Then, it computes the minimum cost plan for changing thestate of rtActiv to be a node in Nod.

If associated plans were not found, the action a would bediscarded for expansion. Otherwise, the successor functionwill return a planning state

�PlS0c

�resulting from executing

the associated plans at PlSc. Notice that order, which isrelated to action dependencies, matter. Overall, the suc-cessor function of a state PlSc returns a set of tuples,where each tuple is composed of: a possibly empty orderedset of associated plans; a possible action; and the result-ing state. PLAN-EA then chooses to expand the successornode having minimum cost f , which is the sum of (i) thepast path-cost g, computed as the number of actions toreach node PlSs from the root node, and (ii) an admissi-ble heuristic h.rtAg, PlSs/, which computes a lower boundof the actions required to reach the goal from PlSs. Itdoes so by relaxing the constraints imposed by the socialmodel specification and runtime properties. Specifically, honly considers a subset of the action dependencies that aretaken into account in the actual planning process. Thus, aswe briefly explain below, most action dependencies han-dled by algorithms 1, 2, and 3 are also considered byh. Nevertheless, it is performed in a far less costly way:applying admissible ‘rules of thumb’ and without invokingPLAN-EA.

Initially, h checks whether the rtAg is located at the sameactivity than the goal (so it can be reached by uttering

illocutions). If it is the case, h selects from Spec and PlSs,all illocutions that are related to rtAg’s role in this activ-ity. For each illocution, h aggregates the cost of uttering it(1) and, if applicable, a lower bound of the cost of addingrequired participants to the activity. Otherwise, if rtAg isnot located at the goal location, then rtAg should (i) exit itscurrent location and (ii) enter the goal location. h estimatesthe cost for both movements, considering also if currentprotocol states allow the participant to exit/enter the cor-responding activities. Additionally, h checks if the actionexecution will reach the goal, because we can consider that,if it is not the case, at least an additional action will berequired.

Finally, as for standard A�, the search will continue untilthe goal is reached or no more nodes can be expanded. Iffound, PLAN-EA will return the plan (i.e. a sequence ofactions from the root to the goal).

4.1. v-mWater Planning Example

Figure 3 illustrates a planning process PLAN-EA(Spec,Mary, PlSc) that considers the static specification (Spec)from Figure 1 and the runtime properties (Sc) depicted atthe top of Figure 3. In particular, Mary enacts a sellerrole and is located at transition t1, and her goal is toregister a water right; the Waiting&Info activity is open(n1 in Spec), and Registration is not created. The root nodePlS0 (in solid black) is initialised to Sc and the succes-sor function considers two seller actions for Mary. First,a1 D enter.s, t1, Waiting&Info/ will directly lead to statePlS1, because the current state of the activity protocol isn1, and sellers can enter at such node. Second, a2 Denter.s, t1, Registration/ cannot be performed, since Maryneeds to wait for a market facilitator to create and openit. As a consequence, function ASCREATE in Algorithm 2looks for plans for InfoMngr and RegMngr and returns pl0

in Figure 3 as the plan for creating the Registration activity.Afterwards, function ASOPEN in Algorithm 3 finds planpl000 for RegMngr to open the activity. Together, associatedplans pl0 and pl000 transit current state to PlS0001 , where a2can now lead to successor state PlS2. Hereafter, becausePlS2 has lower cost † (f D 5) than the one ‡ of PlS1(f D 6), the planning expands PlS2 and continues untilit reaches the goal node PlS6 (diamond shape indicatesanother participant, mf , did the action) and returns plan plin Figure 3.

†g.PlS2/ D 3 because the associated plans add two actionsto the one performed by Mary. h.PlS2/ D 2 because, althoughMary is at the Registration, none of her possible actions in PlS2

do lead to the goal state.‡g.PlS1/ D 1 because only one action (enter) has been exe-cuted. h.PlS1/ D 5 because Mary is not where the goal taskis performed. Thus, she has, at least, to exit Waiting&Info andenter Registration, which must first be created and opened.Once inside the activity, at least one illocution should be uttered,because the goal is not just to enter the activity.

Comp. Anim. Virtual Worlds 2014; 25:495–504 © 2014 John Wiley & Sons, Ltd. 499DOI: 10.1002/cav

Page 6: Assistant agents to advice users in hybrid structured 3D virtual environments

Assistant agents to advice users in hybrid structured 3D VE P. Almajano et al.

Figure 3. Example of PLAN-EA returning plan pl with associatedplans pl0 and pl000.

4.2. Plan Presentation to Users

Computed plans are sequences of actions meant for soft-ware agents. PAs facilitate the interaction with humanusers by presenting them in natural language, so thathuman user, Mary, can easily interpret the plan inFigure 2.

Action translation requires the designer to specify atemplate. For example, the register illocution is describedas ‘$sender$, ask for registering a water right to$receiver$’,where $sender$ and $receiver$ are substitutedby the name of the actual participants, as last sentencein Figure 2 reads. Moreover, if the receiver is the user,then a template is also used to generate sentences suchas ‘Information Manager will provide information aboutlast transactions to Mary Smith’ (see third sentence inFigure 2).

5. SYSTEM EVALUATION

We evaluated the assistance service§ by (i) fulfillinga usability test that follows the formative evaluationmethodology and (ii) comparing the obtained results withour previous study [3], where the users performed the sametask in v-mWater but without assistance.

5.1. Test Goals and Research Questions

We conducted a user evaluation whose goal was twofold.First, we aimed to evaluate our assistance in terms of its(i) effectiveness, if it helps the users to actually performthe task; (ii) efficiency, if it reduces the effort (in termsof number of user’s actions and cognitive load) requiredto conduct the task; and (iii) users’ satisfaction: theiropinions, feelings and experience. Second, we also aimat identifying the errors/problems users make/encounterwhen using the assistance.

Having these goals in mind, we addressed the followingresearch questions: RQ1: Assistance helpfulness. At whatstage of task completion was the help requested? Wasthe provided advice useful for the user to complete thetask? RQ2: User-personal assistant (PA) interaction. Is theassistance easy to request? How easy and pleasant is theinteraction with the PA? RQ3: Plan tracking. What obsta-cles do users encounter when following the plan? Is itclearly explained? Is it detailed enough to complete thetask successfully and in a seamless way? RQ4: Task com-pletion. How many users do complete the task? How dothey perceive it?

For this usability test, users were asked to perform arather complex task: to register a water right in v-mWaterat a price that depends on previous market transactions.It implies four subtasks:(i) to understand the task (theyare required to visit two rooms in a specific order); (ii) toobtain particular information about the market prices atthe Waiting&Info room (this subtask can be accomplishedby asking the Information Manager bot or by reading theinformation panel); (iii) to come up with the required reg-istration price, which has to be 5e higher than the priceof the most recent transaction; and (iv) to register thewater right by interacting with the Registration bot at theRegistration room. Whenever needed, users can ask forassistance to their PA.

5.2. Participants and Methodology

We recruited 14 participants for our experiment. Table Ishows details on their age, gender, computer skills (‘basic’stands for users of limited computer functionalities and‘advanced’ for computer professionals such as program-mers) and VE experience (‘none’/‘high’ describe userswho have never/often used a VE).

§We encourage the reader to watch http://youtu.be/VOQ9DavaqNA

500 Comp. Anim. Virtual Worlds 2014; 25:495–504 © 2014 John Wiley & Sons, Ltd.DOI: 10.1002/cav

Page 7: Assistant agents to advice users in hybrid structured 3D virtual environments

P. Almajano et al. Assistant agents to advice users in hybrid structured 3D VE

Table I. List of participants’ characteristics.

Name Age Gender PC exp VE exp

P1 23 Female Advanced NoneP2 24 Male Advanced HighP3 26 Female Advanced HighP4 27 Male Advanced HighP5 27 Male Advanced HighP6 27 Male Advanced HighP7 29 Female Advanced NoneP8 32 Male Basic NoneP9 32 Male Basic NoneP10 32 Male Advanced HighP11 41 Male Advanced HighP12 42 Male Basic HighP13 53 Male Advanced NoneP14 66 Female Basic None

Because we are mostly interested in finding relevantqualitative and quantitative data, our usability test is sum-mative and follows the formative evaluation. The tests tookplace at users’ locations: 30% of the participants did thetest at their home and the rest at their workplace, on aseparate room. The equipment consisted in one portablecomputer. It had the overall system installed: OCMAS exe-cution infrastructure [12], OpenSimulator VW server anda VW client. It also recorded user interactions and sound.All participants were requested to perform the aforemen-tioned task by telling them the following: ‘act as a seller,and register a water right for a price which is 5e higherthan the price of the last transaction done’.

A moderator guided the test along four different phases:(1) Pre-test interview: the moderator welcomed the user,briefly introduced the test and asked the user about heror his experience with similar VEs. (2) Training: themoderator taught the user to move in a 3D demo VE aswell as to interact with objects, avatars, bots (whose ’spe-cial’ appearance, i.e. bold and coloured skin, was madenoticeable) and her or his PA. This training part wasmostly fully guided, except at the end, when the user couldfreely roam and interact in the demo scenario. (3) Test:the user performed the assigned task without receiving anyguidance (unless s/he ran out of resources). Meanwhile,the moderator encouraged the user to think-aloud (i.e. todescribe her or his actions and thoughts) while performingthe test. (4) Post-test satisfaction survey: the moderatorgave the user a survey with qualitative (open-ended) andquantitative (close-ended) questions regarding v-mWaterand the assistance provided.

5.3. Results and Discussion

In this section, we show and discuss results obtained afterwe analysed data gathered during the test, i.e desktop andvoice recordings, moderator notes, users’ comments andpost-test satisfaction surveys.

Table II. Post-test questionnaire.

Question Brief description

Q1 Info gathering (panel/bot)Q2 Human-bot interactionQ3 Bot visual distinctionQ4 Dialogue-based bot communicationQ5 Overall system opinionQ6 Assistance usefulnessQ7 PA interactionQ8 Advice Understandingopen Q User’s comments

Table II summarises the eight questions in the post-testsurvey, and Figure 4 depicts users’ answers. There, x-axisshows the questions, and the y-axis shows average valuesof answers considering a 5-point Likert scale. This scaleprovides five different alternatives in terms of applica-tion successfulness (‘very bad’/‘bad’/‘fair’/‘good’/‘verygood’), where ‘very bad’ corresponds to 1, and ‘very good’to 5.

We collected data on general issues about the 3D envi-ronment in questions Q1–Q5, results are similar to thosewe obtained in the evaluation without assistance [3]. It isworth noting that the new group of questions related toassistance (Q6–Q8) have scores over 3.8. We extracted anumber of relevant aspects of assistance in v-mWater fromquestionnaire qualitative measures as well as from userdebriefings with the evaluation team. Generally, users likethe way that the assistance was provided, and how it helpstask accomplishment. The assistance clearly guided themto perform the task and corrected users behaviour whenthey deviated from the proper one. The overall opinion ofthe system was positive.

Usability criteria, such as effectiveness, efficiency, errorsand satisfaction have been analysed answering the researchquestions introduced in Section 5.1.RQ1: Assistance helpfulness. The assistance wasvoluntarily requested by the 93% of the testers: (i) at thebeginning of the task (six users); (ii) in the middle of thetask (four users), to check if they were doing it properly;and (iii) when they were trying to register without gettingprice information (three users). Therefore, the assistancewas helpful to all users at some time during the task. Thisfact is reinforced by the answers to Q6, which has anaverage value of 4.2.RQ2: User-PA interaction. Along the test, all users inter-acted with the PA in a seamless way: they managed toask for the correct advice and recognised the receivedplan. Thus, user-PA interaction was satisfactory (as Q7 alsoindicates with a value over 4).RQ3: Plan tracking. All users who requested the planunderstood its structure and followed it without errors. Ingeneral, the advice was highly comprehensible. Relatedquestion Q8 has a value of 3.8.RQ4: Task Completion. The difference in number ofactions with assistance (average of 7, � D 2.3) and withoutit (average of 10.8, � D 3.4) has proven to be significant

Comp. Anim. Virtual Worlds 2014; 25:495–504 © 2014 John Wiley & Sons, Ltd. 501DOI: 10.1002/cav

Page 8: Assistant agents to advice users in hybrid structured 3D virtual environments

Assistant agents to advice users in hybrid structured 3D VE P. Almajano et al.

Figure 4. Post-test questionnaire average results and standard deviations.

by a one tailed unequal variance t-test with a p-value of0.004 (p-value < 0.05). In order to successfully completethe given task, users had to perform a minimum of fiveactions (see Mary’s actions on the left of Figure 2). If weanalyse these averages with and without assistance withrespect to this minimum, they represent a 140% and 216%,respectively, so that assistance provides a 76% reduction.Assuming we are considering the same type of actions (i.e.the ones defined in the social model), we take the num-ber of actions as a measure of efficiency. Assisted userswent more directly to the goal and, thus, had less cognitiveload than non-assisted ones. Overall, assistance reducesperceived task complexity, and this fact is confirmed bythe lower percentage of failures with assistance (7%) withrespect to those without it (23%).

6. CONCLUSION

We propose PA agents for advising users to perform com-plex tasks in hybrid structured 3D VE. These VE arehybrid because participants can be human and SW, andstructured because most interactions are regulated by anOCMAS. A PA provides its assisted user with a plan, asequence of (OCMAS compliant) user actions, that willbe extended with other participants’ actions wheneverneeded to accomplish user’s task. To do so, we proposean extension of A�, namely PLAN-EA. Evaluation resultsindicate that assistance impacts positively in usability mea-sures of efficiency, efficacy and satisfaction. A comparativeanalysis of the number of user actions with (7) and without(10.8) assistance showed a significant difference. More-over, with assistance, 93% of users completed the task,compared to 77% of users without assistance. Finally,users liked the way assistance was provided, and how itfacilitated task completion.

We are currently studying PAs in the smart microgridsdomain [14]. As future work, we plan to improve theway plans are presented to users, to make PAs proactive,and to extend their capabilities with both justification andestimation services.

ACKNOWLEDGEMENTS

Work funded by Spanish CSD2007-0022, TIN2011-24220and TIN2012-38876-C02-02 research projects.

REFERENCES

1. Ferber J, Gutknecht O, Michel F. From agents toorganizations: an organizational view of multi-agentsystems. In Agent-Oriented Software EngineeringIV. Lecture Notes in Computer Science, Vol. 2935.Springer: Berlin Heidelberg, 2004; 214–230.

2. Almajano P, Lopez-Sanchez M, Rodriguez I. An assis-tance infrastructure to inform agents for decision sup-port in open mas. In ITMAS’12, Valencia, Spain, 2012;93–106.

3. Almajano P, Mayas E, Rodriguez I, Lopez-Sanchez M,Puig A. Structuring interactions in a hybrid virtualenvironment: infrastructure & usability. In GRAPP’13,Barcelona, Spain, 2013; 288–297.

4. Dong R, McCarthy K, O’Mahony M, Schaal M,Smyth B. Towards an intelligent reviewer’s assistant:recommending topics to help users to write betterproduct reviews. In IUI’12, Lisbon, Portugal, 2012;159–168, ACM.

5. Kumar S, Kunjithapatham A, Sheshagiri M, Finin T,Joshi A, Peng Y, Cost RS. A personal agent applicationfor the semantic web. In AAAI 2002 Fall SymposiumSeries, North Falmouth, Massachusetts, November2002; 30–37.

6. Lujak M, Billhardt H. Coordinating emergencymedical assistance. In Agreement Technologies. Law,Governance and Technology Series, Vol. 8. Springer:Netherlands, 2013; 597–609.

7. Faulring A, Myers B, Mohnkern K, Schmerl B,Steinfeld A, Zimmerman J, Smailagic A, Hansen J,Siewiorek D. Agent-assisted task management that

502 Comp. Anim. Virtual Worlds 2014; 25:495–504 © 2014 John Wiley & Sons, Ltd.DOI: 10.1002/cav

Page 9: Assistant agents to advice users in hybrid structured 3D virtual environments

P. Almajano et al. Assistant agents to advice users in hybrid structured 3D VE

reduces email overload. In IUI’10, Hong Kong, China,2010; 61–70, ACM.

8. Chalupsky H, Gil Y, Knoblock CA, Lerman K,Oh J, Pynadath DV, Russ TA, Tambe M. Electric elves:applying agent technology to support human organiza-tions. In IAAI’01, Seattle, Washington, 2001; 51–58,AAAI Press.

9. Oh J, Meneguzzi F, Sycara K, Norman TJ. Prognos-tic normative reasoning. Engineering Applications ofArtificial Intelligence 2013; 26(2): 863–872.

10. van Dijk E. MAG, op den Akker HJA, Nijholt A,Zwiers J. Navigation assistance in virtual worlds.Informing Science, Special Series on Community Infor-matics 2003; 6: 115–125.

11. Chittaro L, Ieronutti L, Ranon R. Navigating 3Dvirtual environments by following embodied agents:a proposal and its informal evaluation on a virtualmuseum application. PsychNology Journal 2004; 2(1):24–42.

12. Esteva M, Rosell B, Rodríguez-Aguilar JA,Arcos JL. AMELI: an agent-based middleware forelectronic institutions. In AAMAS’04, New York, NY,USA, 2004; 236–243.

13. Hart PE, Nilsson NJ, Raphael B. A formal basis for theheuristic determination of minimum cost paths. IEEETransactions on Systems Science and Cybernetics1968; 4(2): 100–107.

14. Bourazeri A, Pitt J, Almajano P, Rodriguez I,Lopez-Sanchez M. Meet the meter: visualising smart-grids using self-organising electronic institutions andserious games. In SASO’12, Lyon, France, 2012;145–150.

AUTHORS’ BIOGRAPHIES

Pablo Almajano studied ComputerScience in the Universidad Ponti-ficia de Comillas (Madrid, Spain)and Artificial Intelligence and Com-puter Vision master, in the UniversitatAutonoma de Barcelona together withthe Artificial Intelligence ResearchInstitute (IIIA,CSIC) both located inBarcelona, Spain.

He is a PhD candidate and part-time lecturer in the AppliedMathematics Department in the Universitat de Barcelona(Spain). His research interests focus on decision support inhybrid (populated by humans and agents) multi-agent sys-tems, human-agent interactions and 3D normative virtualworlds. He has participated in several research projects andhas a variety of scientific papers published in indexed jour-nals, books and international conferences. He also has beenreviewer in international conferences.

Maite Lopez-Sanchez is a ComputerScientist that holds, since 1999, aPhD on Artificial Intelligence from theArtificial Intelligence Research Insti-tute (IIIA-CSIC, http://www.iiia.csic.es/). As a pre-doc, she was a visitorresearcher at the University of South-ern California (USC, http://www.usc.edu/), and as a post-doc, she was

co-founder of iSOCO (intelligent Software Componentshttp://www.isoco.com/), a spin-off company from the IIIA.At iSOCO. She was a research manager and was in chargeof a number of European and national research projects.Currently, she is an Associate Professor (TU) at the Uni-versity of Barcelona (UB, http://www.ub.edu). She is thecoordinator at UB of the interuniversitary master on Arti-ficial Intelligence and some national research projects.Overall, she has published about 100 scientific publica-tions in indexed journals, books and ranked internationalconferences, and she has supervised both PhD and mas-ter students in the area Normative Multi-Agent Systems. Inaddition to research papers, she has also coauthored severalvideo and demo contributions (some of which have beenawarded).

Inmaculada Rodriguez studied com-puter science in the University ofGranada (Spain) and obtained her PhDin the University of Alcala (Spain)in 2004. She is associate professor inthe Applied Mathematics Departmentin the University of Barcelona whereshe is member of WAI (Volume Visu-alization and Artificial Intelligence)

research group. She has worked in the computer animationof virtual humans, and currently, her major research focuslies on the integration of both artificial intelligence and vir-tual worlds technologies, serious games and gamification.She has participated in several national and internationalprojects since 1998. She has published over 50 researchpapers in international conferences and journals, and she isa member of program committee in different internationalconferences and journals in the fields of Computer Graph-ics and User Interfaces. Additionally, she has supervisedboth PhD and master students in the research line of Nor-mative Virtual Worlds. Web page: http://www.maia.ub.es/~inma/

Tomas Trescak holds a PhD titlein Computer Science with speciali-sation in artificial intelligence fromArtificial Intelligence Research Insti-tute, Barcelona, Spain (IIIA) of theSpanish Research Council (CSIC).Since May 2013, he is a PostdoctoralResearcher Fellow in Digital Human-ities, in the University of Western

Sydney, Australia.

Comp. Anim. Virtual Worlds 2014; 25:495–504 © 2014 John Wiley & Sons, Ltd. 503DOI: 10.1002/cav

Page 10: Assistant agents to advice users in hybrid structured 3D virtual environments

Assistant agents to advice users in hybrid structured 3D VE P. Almajano et al.

The topics of his research concern interactive virtualworlds and specifically normative virtual worlds, intel-ligent virtual agents, crowd simulation and computa-tional design techniques, such as shape grammars. He’smain contribution is in facilitation of creation and execu-tion of interactive normative 3D virtual worlds and their

subsequent application to the fields of social, culturaland historical agent-based simulation. During his research,he has developed several techniques and methods andimplemented them in the set of open-source tools, usedworld-wide. These works were published in acclaimedinternational conferences and journals.

504 Comp. Anim. Virtual Worlds 2014; 25:495–504 © 2014 John Wiley & Sons, Ltd.DOI: 10.1002/cav