An Event Metric and an Episode Metric for a Virtual Guide

Felix Rabe and Ipke Wachsmuth

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

Keywords:

Episodic Memory, Event Metric, Episode Metric, Virtual Agent, Guidance.

Abstract:

In this paper we introduce a metric to compare events and episodes in the episodic memory system of a virtual

agent. The agent, a virtual tour guide based on a belief – desire – intention cognitive architecture, uses his

memories to improve the walks around a virtual city. The guide’s past experiences are memorized as events

and organized into episodes. Each event is indexed along six dimensions and is comparable on each dimension

with a distinct distance function. This is then utilized to measure the similarity between episodes.

1 INTRODUCTION

Past experiences inﬂuence all of our actions and let

us have an expectation of what might happen next. In

interaction we rely on past episodes with other per-

sons, we improve our behavior based on what we ex-

perience and store in our episodic memory. Since

all of our actions are inﬂuenced by our past experi-

ences, it is important also for a virtual agent to have an

episodic memory, so his behavior in interaction with

humans is improved.

Our work is centered around a virtual humanoid

agent Max, cf. (Leßmann et al., 2006; Becker et al.,

2006) that has already a lot of skills and is based on

a belief – desire – intention (BDI) cognitive archi-

tecture. In (Rabe and Wachsmuth, 2012) we intro-

duced a cognitively motivated episodic memory for

our agent and a virtual guide scenario, where episodic

memory is employed. The memories are episodes of

events, that means past episodes (guided tours) con-

sist of multiple events (e.g. explaining a sight). Each

event can be accessed via six different indices: time,

space, protagonists, intention, causality and emotion.

In this paper we introduce a metric with several

distance functions to compare events among each

other and measure similarity between episodes.

2 RELATED WORK

Our work on constructing a memory system for a vir-

tual guide is inﬂuenced by theories on episodic mem-

ory, as well as on event-indexing and event segmenta-

tion. Episodic memory, as deﬁned in (Tulving, 1972),

deals with temporally dated episodes or events, and

temporal-spatial relations among these events. Every

“item in episodic memory” (Tulving) is a more or less

faithful record of a person’s experience of an occur-

rence.

The Event-Indexing Model (Zwaan et al., 1995)

describes how readers of short stories construct a

model of the situation in the text. As readers un-

derstand what is happening in the story they update

the model along ﬁve indices: Time, Space, Causal-

ity, Intentionality and Protagonists. These dimen-

sions store answers to the questions of what happened

when, where, why and how, and who was involved.

In Event Segmentation Theory, an event is de-

ﬁned as “a segment of time at a given location that

is conceived by an observer to have a beginning and

an end” (Zacks and Tversky, 2001). (Allen et al.,

2008) propose that emotion is an important contex-

tual cue for episodic memory and provide evidence

that cognition is either moderated or mediated by ba-

sic affective processing. Another related ﬁeld is Ex-

perience Management. The application makes use of

several distance functions to provide matching solu-

tions (Bergmann, 2002).

In earlier work (Rabe and Wachsmuth, 2012) we

discussed several other applications of agents em-

ployed as guides and computational episodic mem-

ory systems. Our recent work is somewhat related to

case-based reasoning (Kolodner, 1993; Aamodt and

Plaza, 1994), except there is no generalization nec-

essary for episodic memory. Concerning the event

metric, (Tecuci and Porter, 2007) propose generic

episodes with three dimensions: context, contents and

outcome. They also provide a semantic similarity

543

Rabe F. and Wachsmuth I..

An Event Metric and an Episode Metric for a Virtual Guide.

DOI: 10.5220/0004262605430546

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 543-546

ISBN: 978-989-8565-39-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

measure (Tecuci and Porter, 2009). Nuxoll and Laird

also have evaluated a variety of metrics for deter-

mining match scores on episode retrival (Nuxoll and

Laird, 2012). Their implementation is similar to our

approach: They calculate a match score based on the

number of elements a cue and memory have in com-

mon and the activation level of the cue that matches

the memory.

3 EVENT INDICES

We deﬁne every observable occurrence as an event,

in contrast to common language, where especially ex-

traordinary occurrences are called event. Second, we

follow the deﬁnition of Zacks and Tversky (Zacks and

Tversky, 2001), that an event is a segment of time with

a beginning and an end. Third, we are only consider-

ing events of equal level, that means we do not con-

sider events that contain other events. In our approach

events are not organized in partonomic event hierar-

chies, but in episodes. Fourth, we index events along

ﬁve dimensions (Time, Space, Protagonists, Inten-

tionality and Causality), according to Zwaan’s event-

indexing model. Fifth, we add emotion as sixth di-

mension. Figure 1 shows how episodes group events

together and how events are conceptualized.

Ep. 1

E1 E5

Ep. 2

E8 E9

E10

E11

Ep. 3

...

E15

E16

E17

E18

Event

Emotion

Causality

Intention

Protagonist

Space

Time

Figure 1: How episodes and events are conceptualized:

Events are grouped into episodes. From the outside of

episodes only events with a strong emotional impact are vis-

ible, 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 (Rabe and Wachsmuth, 2012).

Independent of the six dimensions a memorized

event can contain information related to the situation

that it represents, we call that Payload. But in contrast

to the indices the additional information can not be

accessed directly.

Time is the recorded system time at the moment the

event begins, and at the moment the event ends. This

information is used to get the duration of the event,

which is also stored. Further possible applications are

to get time of the day the event occurred, and to order

events in time.

Space is represented in the virtual world coordinates,

but indexed along named places to which the coor-

dinates have been mapped. Similar to humans, who

normally do not tend to memorize places in GPS coor-

dinates, the agent memorizes the name of the place an

event happened. All places are also represented using

a undirected weighted location graph, that means the

agents knows which places are connected and how far

they are apart. Also the agent’s knowledge about the

places is accessible by names and not by coordinates.

Nevertheless we keep the coordinates for further anal-

ysis.

Protagonists stores named representation of the in-

dividuals present during an event. This can be the

agent (‘I’) himself and any known and named visi-

tors. This enables the agent to remember earlier oc-

currences with the same visitors. In our current setup

the agent can recognize only one visitor in the Virtual

Environment.

Intention represents the intention the agent has dur-

ing the current event. Since the agent is based on

a BDI cognitive architecture the agent next actions

are based on his intention. In our BDI-architecture

it is the name of the plan the agent is currently pursu-

ing. All of names the agent’s plans are composed of

two parts, an action and a token separated by a dash,

e.g. smalltalk-greet, explain-navigation,

follow-visitor, show-city hall. All actions can

be categorized into either watching, guiding, explain-

ing, or small-talking.

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

either a named percept of the agent (e.g. a request

stated by the visitor to show a certain building), or a

named action the agent performed (e.g. completely

leading the way to a certain building). Similary to

the intention names causes’ names consist of three

parts, an actor, an action, and a token each separated

by a dash, e.g. visitor-focused-jazzclub,

guide-proposed-city overview,

visitor-reached-marketplace.

Emotion contains the current emotional state of the

agent (in PAD space) and may also contain the emo-

tional impulse the agent receives. It is our addition to

the original event-indexing model, since the emotion

may include clues for different things: If the guide re-

ceives good feedback, his emotional system rewards

him and he is happy. If he receives negative feed-

back, his emotional system dampens his mood and he

is sad. In further similar occurrences the guide will

tend to redo the things that made him happy.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

544

4 EVENT METRIC

Our event metric is a metric space, an ordered pair

(M,d) where M is a set of events and d is a metric on

M, i.e., a distance function

d : M ×M → R (1)

We designed the distance function to not produce

results greater than 1. All distance functions are equal

weighted. This means we added the following con-

straint:

d : M ×M → [0,1] (2)

Now we deﬁne I as a set of all indices:

I = {t,s, p, i, c, e} (3)

Consider an event e, which is indexed along the

six dimensions of time, space, protagonists, intention-

ality, causality and emotion:

e = (e

) (4)

The naive distance between two events e and f is

computed as the sum of the discrete distances of the

individual indices divided by the cardinality of I:

(e, f ) =

|I|

∑

i∈I

, f

) (5)

Here we use the discrete distance funticon d

that means we only compare if the indices are equal

or not in each dimension:

(x,y) =



0 if x = y

1 if x 6= y

(6)

The minimal naive distance is 0 (all indices are

alike) and the maximal is 1 (none are alike).

To enhance the event metric, we deﬁne distinct

distances for each index which gives us the overall

index-distinct distance:

(x,y) =

) + d

)

) + d

)

(7)

We compare events on the time index using the

duration of the events. To normalize the distance we

divide the duration difference by the longest duration

of all events. This gives us the time-durations dis-

tance:

(x,y) =

−y

max

duration

(8)

We compare events on the space index using a

waypoint distance. Due to the scenario the streets

and places form an undirected weighted graph. We

calculate the shortest path using Dijkstra’s algorithm

and divide the length of the result by the length of the

longest path of our scenario:

(x,y) =

Dijkstra

)

max

distance

(9)

For a different scenario it might be applicable to

use the Euclidean distance instead.

We compare events on the protagonist index us-

ing the discrete distance function as protagonist dis-

tance. In future work we think about incorporating

a relationship based distance using work from (Mat-

tar and Wachsmuth, 2012) who are building a person

memory for the virtual agent.

The intention distance is the sum of a category

based distance for the action part of the intention and

the discrete distance for the token part. The category

based distance is an extended discrete distance, where

if an action x is not equal to y but in the same action

category C as y the distance is smaller than if x and y

would be in different action categories:

(x,y) =







0 if x = y

0.5 if x 6= y ∧x,y ∈C

1 if x 6= y ∧x ∈C ∧y /∈C

(10)

With this the intention distance is:

action

) + d

token

))

(11)

We compare events on the causality index using a

category based distance for the action part of the in-

tention and the discrete distance for the actor and to-

ken parts. Similar to the intention distance the causal-

ity distance is:

actor

) + d

action

token

))

(12)

Finally, the agent’s emotional state is mapped

to a 2.5 dimensional representation of the PAD

space (Becker-Asano and Wachsmuth, 2010). The

maximal expansion from (−100, −100, −100) to

(100,100,100) is max

PAD

= 200

√

3 ≈ 346.41. We

use this to calculate a normalized Euclidean distance,

the mood distance:

kx −yk

max

PAD

(13)

5 EPISODE METRIC

To determine which episodes are alike we have a

look at how may events of the episodes to compare

are similar. Therefore we deﬁne an episode metric

(M,d) (similar to the event metric) where M is a set

AnEventMetricandanEpisodeMetricforaVirtualGuide

545

of episodes which consist of events. Considering two

episodes E and F we deﬁne the episode distance as

(E,F) = 1 −

∑

e∈E

∑

f ∈F

(e, f )

|E|·|F|

, (14)

where d

is a discrete distance which is 0 if the re-

sult of the index-distinct distance is below a certain

boundary b:

(e, f ) =



0 if d

(e, f ) < b

1 if d

(e, f ) ≥ b

(15)

This means we compare every event e of episode

E to every event f in episode F and count the num-

ber of matches. This number is then divided by the

product of the number of all elements in E and F and

then subtracted from 1. A smaller result means that

the episodes are more similar.

Note that depending on how b is selected it can

be sufﬁcient if only one or two indices of the events

to compare are similar enough. If e.g. b = 1 −

two

events would match if two indices are alike.

6 CONCLUSIONS & OUTLOOK

We have introduced an event metric and an episode

metric which our virtual guide employs to select

memories matching to the current situation. He uti-

lizes this memories in his decision process what to do

next.

Next steps of our work embrace collecting more

data, that means that our agent has to give many more

tours to accumulate rich memories. With that we plan

to evaluate the performance of the memory system

and measure the quality of the chosen actions.

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

Aamodt, A. and Plaza, E. (1994). Case-based reasoning:

Foundational issues, methodological variations, and

system approaches. AI Communications, 7(1):39–59.

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.

Becker, C., Leßmann, N., Kopp, S., and Wachsmuth, I.

(2006). Connecting feelings and thoughts - modeling

the interaction of emotion and cognition in embodied

agents. In Fum, D., Del Missier, F., and Stocco, A.,

editors, Proceedings of ICCM06, pages 32–37, Tri-

este, Italy. Edizioni Goliardiche.

Becker-Asano, C. and Wachsmuth, I. (2010). Affective

computing with primary and secondary emotions in

a virtual human. Autonomous Agents and Multi-Agent

Systems, 20:32–49.

Bergmann, R. (2002). Experience Management, volume

2432 of LNAI. Springer, Berlin, Germany.

Kolodner, J. L. (1993). Case-Based Reasoning. Morgan

Kaufmann, San Mateo, CA.

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, Germany.

Mattar, N. and Wachsmuth, I. (2012). Who are you? on the

acquisition of information about people for an agent

that remembers. In Filipe, J. and Fred, A., editors,

Proceedings of ICAART 2012, volume 2, pages 98–

105, Vilamoura, Portugal. SciTePress.

Nuxoll, A. M. and Laird, J. E. (2012). Enhancing intelli-

gent agents with episodic memory. Cognitive Systems

Research, 17–18:34–48.

Rabe, F. and Wachsmuth, I. (2012). Cognitively moti-

vated episodic memory for a virtual guide. In Fil-

ipe, J. and Fred, A., editors, Proceedings of ICAART

2012, volume 1, pages 524–527, Vilamoura, Portugal.

SciTePress.

Tecuci, D. and Porter, B. (2009). Memory based goal

schema recognition. In Lane, H. C. and Guesgen,

H. W., editors, Proceedings of FLAIRS-09, Menlo

Park, CA. AAAI Press.

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, California. The AAAI Press.

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.

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.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

546