COGNITIVELY MOTIVATED EPISODIC MEMORY

FOR A VIRTUAL GUIDE

Felix Rabe and Ipke Wachsmuth

Artiﬁcial Intelligence Group, Faculty of Technology, Bielefeld University, Bielefeld, Germany

Keywords:

Episodic memory, Virtual reality, Guidance, Event index, Embodied conversational agent.

Abstract:

This paper describes how the guiding capabilities of a virtual agent with a belief – desire – intention cognitive

architecture can be enhanced by adding an episodic memory. We describe how the theories of episodic memory

and event segmentation can be applied to the architecture of our virtual agent Max, and how to create an index

according to the event indexing model. Having memories of past experiences will enable our agent to have

improved plans how to react in future interaction.

1 INTRODUCTION

Past experiences inﬂuence all of our actions and let

us have an expectation of what might happen next, so

that we are able to plan well. Especially in interac-

tion we rely on past episodes with other persons, we

improve our behavior based on what we experience

and store in our episodic memory. To improve the be-

havior of virtual agents in interaction with humans the

agents need episodic memory, too.

In the past we have introduced the virtual hu-

manoid agent Max (Leßmann et al., 2006) who al-

ready has lots of skills and is based on a belief –

desire – intention (BDI) cognitive architecture. He

has proven a helpful interaction partner, e.g. in assist-

ing a human in construction tasks. In a different sce-

nario Max is “working” as a museums guide, where

he stands in a museum, conducts small talk with vis-

itors and explains sights. But up to now Max has no

memory of his own actions and his experiences. He

can only memorize some facts about persons visiting

him. With the cognitively motivated episodic memory

this is to be changed.

We have conceived a virtual guide scenario, where

Max can utilize his new skill in interaction. We picked

a virtual reality (VR) scenario, since VR is where hu-

mans and machines meet and can interact at eye level.

The episodic memory system will be especially useful

when the agent is guiding a visitor through a virtual

environment. The main focus will be on the agent

memorizing his own actions, the interaction with the

human visitor, and the improvement of the guidance

by assisting the human.

2 RELATED WORK

2.1 Psychological Background

Endel Tulving suggested that there are two distinct

types of declarative long-term memory: Episodic

and semantic memory. Semantic memory is fac-

tual knowledge about the world. Episodic memory

deals with temporally dated episodes or events, and

temporal-spatial relations among these events. Every

event, or “item in episodic memory”, as it is called by

Tulving, is a more or less faithful record of a person’s

experience of an occurrence, and can include the per-

ceptible properties of that moment (Tulving, 1972).

Another theory we build on is the Event-Indexing

Model (Zwaan et al., 1995). It describes how readers

of short stories construct a model of the situation in

the text. As the readers understand what is happen-

ing in the story they update the model along ﬁve in-

dices: Time, Space, Causality, Intentionality and Pro-

tagonist. These dimensions store the answers to the

questions of what happened when, where, why and

how, and who was involved. Causality is what lead to

the current event and Intentionality are the intentions

the protagonists had during the event. Zwaan and col-

leagues suggested that events can be compared along

these indices: The more indices differ, the more dis-

tinct are the events.

In Zacks and Tversky’s Event Segmentation The-

ory (Zacks and Tversky, 2001), an event is deﬁned

as “a segment of time at a given location that is con-

ceived by an observer to have a beginning and an

end.” Second, events can be organized in partonomic

524

Rabe F. and Wachsmuth I..

COGNITIVELY MOTIVATED EPISODIC MEMORY FOR A VIRTUAL GUIDE.

DOI: 10.5220/0003882205240527

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 524-527

ISBN: 978-989-8425-95-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

hierarchies, which means they may span very long

and very short periods of time. A long event can em-

brace some short events. In further work Zacks et al.

found evidence that when human perception does not

match the internal prediction, an event boundary is

perceived (Zacks et al., 2007).

Allen, Kaut, and Lord propose that emotion is an

important contextual cue for episodic memory and

provide evidence that cognition is either moderated or

mediated by basic affective processing (Allen et al.,

2008).

2.2 Computational Episodic Memory

In computer science, Ho built an autobiographic

memory system for an agent to locate resources en-

countered before and used it in further work to en-

hance virtual characters, so that they were able to talk

about personal past experiences (Ho and Dautenhahn,

2008).

Tecuci and Porter created a generic memory mod-

ule for events, where a generic episode has three di-

mensions: context, contents, and outcome. Context

is the general setting in which an episode happened,

contents is the ordered set of events, that make up the

episode, and outcome is an evaluation of the episodes

effect (Tecuci and Porter, 2007).

Nuxoll and Laird extended a cognitive architec-

ture (Soar) with episodic memory. They showed that

an autonomous agent, a virtual tank, performs better

if it can use episodic memory for reasoning (Nuxoll

and Laird, 2007).

Brom et al. proposed a that virtual characters

need a “full episodic memory” and implemented it

for a non-player character of a computer role-playing

game. The system allows the reconstruction of the

personal story of the character (Brom et al., 2007).

Sieber and Krenn proposed an episodic memory

for a companion to enhance its dialogue capabilities.

The episodes are stored as RDF triples (Sieber and

Krenn, 2010).

2.3 Applications

In the ﬁeld of virtual humanoid agents, there are a few

scenarios where the agent is a guide.

Theune et al. presented a virtual tour guide with

focus on conversational interactions. Their guide

served as direction giving guide in a virtual museum.

They focused on dialogue management and how the

directions can be given with language and gesture

(Theune et al., 2007).

Jan et al. presented a virtual guide in Second Life.

Their agent can be asked about items on an island and

can navigate a user to different spots by walking in

front of him (Jan et al., 2009).

Lim and Aylett built an affective mobile guide,

which has an emotion memory. The stories the guide

tells about the environment are inﬂuenced by emo-

tions he experienced beforehand (Lim et al., 2009).

3 MEMORY MODEL

We will explain the cognitive architecture of Max and

how we extend it with memory (see ﬁg. 1).

Memories

Perceive

Act

Reason

Remember

Emotions

Beliefs Desires

Sensors

Reactive Behavior

Mediator

Deliberative

Behavior

Figure 1: This ﬁgure shows the cognitive architecture of

Max, extended with memory.

In the virtual world Max perceives dialogue input

and other actions of the user, e.g. navigation to a loca-

tion and focusing on a speciﬁc virtual item. Max also

knows where he and the user are.

Due to what Max perceives he reasons and de-

liberatively chooses his actions. The perception pro-

cess also delegates the input to the memory which has

an “unconscious” mechanism that segments the in-

put into events, following the idea stated by Zacks et

al. that event segmentation is a spontaneous outcome.

During reasoning Max can choose to actively remem-

ber episodes and events from his memory. Also there

is the possibility that his memory comes up with an

episode or event that is similar to the current situa-

tion, Max then has a spontaneous inspiration.

3.1 Events and Episodes

In common language event is a board concept. To nar-

COGNITIVELY MOTIVATED EPISODIC MEMORY FOR A VIRTUAL GUIDE

525

row it down for our purposes, ﬁrst of all an event is

an observable occurrence, and for us any observable

occurrence is an event, not only extraordinary occur-

rences. Second, we follow the deﬁnition of Zacks and

Tversky, that an event is a segment of time, that an

event has a beginning and an end. Third, at the mo-

ment we are only considering events of equal level,

that means we don not consider events that contain

other events. So, our events are not (yet) organized

in partonomic event hierarchies, but in episodes. Fig-

ure 2 shows how we organize events and episodes.

Ep. 1

E1 E5

Ep. 2

E8 E9

E10

E11

Ep. 3

...

E15

E16

E17

E18

Event

Emotion

Causality

Intention

Protagonist

Space

Time

Figure 2: This ﬁgure shows how episodes and events are

conceptualized. From the outside of episodes only events

with a strong emotional impact are visible, e.g. event E2.

The enlarged event E11 shows the six indices. Episode Ep.

3 is the current episode to which new events can be added.

Max’s events are indexed according to Zwaan’s

event-indexing model along ﬁve indices. But since

Allen et al. state that emotion is also an important cue

for episodic memory we extended the model with a

sixth dimension.

Time. Time is the recorded system time at the mo-

ment the event begins, at the moment the event

ends. This information can be used to get either

the length of the event and to get time of the day

the event occurred.

Space. Space is represented in the virtual world coor-

dinates. The agent can map these to named places.

Protagonist. Can be the agent (‘I’) himself and any

known and named visitor.

Intention. This is the intention the agent has in the

current event. It is a reference to the plan the agent

is currently pursuing.

Causality. This is what has lead to the event. This

can be either a named action the agent performed,

or a named percept of the agent.

Emotion. This is the current emotional state of the

agent as well as the emotional impulse the agent

receives.

Additional Payload. Furthermore we add additional

“payload” to the event. This may include percep-

tual stimuli of the agent, e.g. a snapshot of what

he heard, saw, and thought.

Events are compared along their indices, the more in-

dices are similar the more two events are alike. This is

useful in ﬁnding related events. If the agent tries to re-

member something he ﬁrst will only be able to access

events with a strong emotional impact, since these are

the events which function as key events to episodes.

Should the agent not be satisﬁed with the recovered

memories, he tries to remember events that may be

hidden in episodes, since the emotional impact is not

that strong. But he should be able to access similar

episodes relatively fast, since the indices are similar

and therefore near.

4 GUIDE SCENARIO

Our scenario is accompanying and guiding a non-

local person through a large and complex virtual en-

vironment: A visitor encounters Max in a CAVE-

like virtual reality of Virtual T

ubingen, a virtual

city realistically modeled after the historic center of

ubingen. It has been originally developed at the Max

Planck Institute for Biological Cybernetics as a natu-

ralistic, controllable environment for investigating hu-

man spatial cognition (van Veen et al., 1998). The

model covers an area of 500x150 m

, it has 15 streets

which are mostly curved and varying in width, and

about 200 houses with different, photorealistic tex-

tures. Max has knowledge about 25 sights and 50

additional waypoints, he uses for his orientation.

After greeting the visitor Max introduces himself

and offers to give a guided tour or to just accompany

the visitor in exploration of the town.

If the visitor chooses to explore the city on his own

Max just accompanies him and watches the visitors

actions. Should Max remember an episode which ﬁts

to some of the visitors actions, Max will ask if he may

provide guidance to additional sights he remembers.

If the visitor chooses to be guided Max has the

goal to give a tour and takes over the steering control.

He remembers a “master” episode where he has given

a tour before, he will visit the same places and give

the same information. If he is not interrupted at all

in the end the current episode will be similar to the

master episode. But since Max can be interrupted at a

certain point he may change the tour due to the wishes

of the visitor. He will then try to adopt with the help

of different tours he has given before and with the

knowledge he has. For example, if the visitor utters

that he is very interested in seeing old churches, then

ICAART 2012 - International Conference on Agents and Artificial Intelligence

526

Max tries to remember an episode where he showed

churches. If he does not remember anything he will

plan a new tour on his own, since he knows that there

are churches. In the end Max then has memorized a

new episode with lots of churches in it.

On the end of a tour Max can be asked to give

a summary of the tour. He will then remember the

whole episode and tell the visitor what places were

visited and which sights have been explained. Fur-

thermore the visitor can ask Max why he has shown

a speciﬁc sight. Due to the Causation and Intention-

ality indices of the events Max can generate adequate

answers.

5 OUTLOOK

We have introduced a cognitively motivated episodic

memory system to be integrated in a BDI-based agent

architecture. Future work is in evaluating the quality

and the beneﬁt of the episodic memories with respect

to Max’s event segmentation capabilities.

One possible way would be to video tape Max

guiding a person through Virtual T

ubingen. After-

wards the recording is played back to a certain num-

ber of probands who segment the video clip into

events. These events than can be compared to the

events Max has segmented on his own and from that

we can see how human like Max event perception is.

Another possibility would be having a human

guiding a visitor through Virtual T

ubingen with Max

only watching the tour without interfering. After-

wards the visitor is asked to retell the tour in form

of events which is compared to Max’s summary.

ACKNOWLEDGEMENTS

This project is supported by the Cognitive Interaction

Technology Excellence Center (CITEC). We grate-

fully acknowledge the MPI for Biological Cybernet-

ics for providing us with Virtual T

ubingen.

REFERENCES

Allen, P. A., Kaut, K. P., and Lord, R. R. (2008). Emo-

tion and episodic memory. In Dere, E., Easton, A.,

Nadel, L., and Huston, J. P., editors, Handbook of

Episodic Memory, pages 115–132. Elsevier, Amster-

dam, Netherlands.

Brom, C., Pe

skova, K., and Lukavsk

y, J. (2007). What

does your actor remember? towards characters with a

full episodic memory. In Cavazza, M. and Donikian,

S., editors, Virtual Storytelling: Using Virtual Real-

ity Technologies for Storytelling (ICVS), LNCS 4871,

pages 89–101, Berlin. Springer.

Ho, W. C. and Dautenhahn, K. (2008). Towards a narrative

mind: The creation of coherent life stories for believ-

able virtual agents. In Prendinger, H., Lester, J., and

Ishizuka, M., editors, Intelligent Virtual Agents, LNAI

5208, pages 59–72, Berlin. Springer.

Jan, D., Roque, A., Leuski, A., Morie, J., and Traum, D. R.

(2009). A virtual tour guide for virtual worlds. In

Ruttkay, Z., Kipp, M., Nijholt, A., and Vilhj

almsson,

H. H., editors, Intelligent Virtual Agents, LNAI 5773,

pages 372–378, Berlin. Springer.

Leßmann, N., Kopp, S., and Wachsmuth, I. (2006). Situ-

ated interaction with a virtual human – perception, ac-

tion, and cognition. In Rickheit, G. and Wachsmuth,

I., editors, Situated Communication, pages 287–323.

Mouton de Gruyter, Berlin.

Lim, M., Aylett, R., Ho, W. C., Enz, S., and Vargas,

P. (2009). A socially-aware memory for compan-

ion agents. In Ruttkay, Z., Kipp, M., Nijholt, A.,

and Vilhj

almsson, H. H., editors, Intelligent Virtual

Agents, LNCS 5773, pages 20–26, Berlin. Springer.

Nuxoll, A. M. and Laird, J. E. (2007). Extending cognitive

architecture with episodic memory. In Holte, R. C.

and Howe, A., editors, Proceedings of the 22

AAAI

Conference on AI, pages 1560–1565, Menlo Park, CA.

AAAI Press.

Sieber, G. and Krenn, B. (2010). Episodic memory for com-

panion dialogue. In Proceedings of the 2010 Work-

shop on Companionable Dialogue Systems, pages 1–

6, Uppsala, Sweden. ACL.

Tecuci, D. G. and Porter, B. W. (2007). A generic memory

module for events. In Wilson, D. C. and Sutcliffe, G.

C. J., editors, Proceedings of FLAIRS-07, pages 152–

157, Menlo Park, CA. AAAI Press.

Theune, M., Hofs, D., and van Kessel, M. (2007). The

virtual guide: A direction giving embodied conversa-

tional agent. In INTERSPEECH-2007, pages 2197–

2200, Antwerp, Belgium.

Tulving, E. (1972). Episodic and semantic memory. In Tul-

ving, E. and Donaldson, W., editors, Organization of

Memory, pages 381–403. Academic Press, New York,

NY.

van Veen, H.-J. A. H. C., Distler, H. K., Braun, S. J., and

ulthoff, H. H. (1998). Navigating through a virtual

city: Using virtual reality technology to study human

action and perception. Future Generation Computer

Systems, 14(3-4):231–242.

Zacks, J. M., Speer, N. K., Swallow, K. M., Braver,

T. S., and Reynolds, J. R. (2007). Event perception:

A mind–brain perspective. Psychological Bulletin,

133(2):273–293.

Zacks, J. M. and Tversky, B. (2001). Event structure in

perception and conception. Psychological Bulletin,

127(1):3–21.

Zwaan, R. A., Langston, M. C., and Graesser, A. C. (1995).

The construction of situation models in narrative com-

prehension: An event-indexing model. Psychological

Science, 6(5):292–297.

COGNITIVELY MOTIVATED EPISODIC MEMORY FOR A VIRTUAL GUIDE

527