From Inter-agent to Intra-agent Representations

Mapping Social Scenarios to Agent-role Descriptions

Giovanni Sileno, Alexander Boer and Tom Van Engers

Leibniz Center for Law, University of Amsterdam, Amsterdam, Netherlands

Keywords:

Agent-roles, Scenario-based Modeling, Social Systems, Institutions, Story Animation, Narratives.

Abstract:

The paper introduces elements of a methodology for the acquisition of descriptions of social scenarios (e.g.

cases) and for their synthesis to agent-based models. It proceeds along three steps. First, the case is analyzed

at signal layer, i.e. the messages exchanged between actors. Second, the signal layer is enriched with implicit

actions, intentions, and conditions necessary for the story to occur. This elicitation is based on elements

provided with the story, common-sense, expert knowledge and direct interaction with the narrator. Third, the

resulting scenario representation is synthesized as agent programs. These scripts correspond to descriptions

of agent-roles observed in that social setting.

1 INTRODUCTION

Research in AI and computer science usually studies

normative systems from an engineering perspective:

the rules of the system are designed assuming that

the behaviour of the system and of its components is

mostly norm-driven. In human societies, however, the

position of policy makers and regulators is completely

different. The target social system exists and behaves

around them, and they are part of it. Moreover, hu-

man behaviour is actually norm-guided. People adapt

to their social environment, both inﬂuencing and be-

ing inﬂuenced by institutions. Multiplicity of institu-

tional conceptualizations and non-compliance (inten-

tional or not) are thus systemic.

We presented in previous works (Boer and van

Engers, 2011b; Boer and van Engers, 2011a; Sileno

et al., 2012) elements of an application framework

based on legal narratives, such as court proceedings,

or scenarios provided by legal experts to make a point

about the implementation or application of the law.

In this, we are investigating a methodology for the

acquisition of computational models of such interpre-

tations of social behaviour. How to represent what

people know about the social system, or, equivalently,

how people (re)act in a social system? Our focus is

These narratives are particularly interesting because

they are produced by the legal system, with the intent of

transmitting - within its current and future components -

relevant social behaviours and associated institutional inter-

pretations.

not on the narrative object, but on the knowledge that

observers and narrators handle, when they observe

social behaviour and generate explanations. The as-

sumption of systemicity is thus not related to the dis-

course, but to a cognitive level.

Reducing the problem to the core, this work in-

vestigates the transformation of a sequence of inter-

agent interactions in intra-agent characterizations, re-

producible in a computational framework. In section

2, we present our case study. We analyze it at a sig-

nal layer, deﬁning the topology and the ﬂow of the

story. In section 3 we show how to enrich the pre-

vious representations with an intentional layer, inte-

grating institutional concepts as well. In section 4 we

provide elements about the transformation of the pre-

vious models into scripts for cognitive agents. Dis-

cussion ends the paper.

2 INTER-AGENT DESCRIPTION

Despite its simplicity, a short story about a sale trans-

action provides a good case of study.

A seller offers a good for a certain amount of

money. A buyer accepts his offer. The buyer

pays the sum. The seller delivers the good.

A successful sale is a fundamental economic transac-

tion. Consequently, what the case describes is a be-

havioural pattern used both in the performance and in

the interpretation of many other scenarios.

622

Sileno G., Boer A. and van Engers T..

From Inter-agent to Intra-agent Representations - Mapping Social Scenarios to Agent-role Descriptions.

DOI: 10.5220/0004909606220631

In Proceedings of the 6th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2014), pages 622-631

ISBN: 978-989-758-015-4

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

seller

buyer

oﬀer a good for a certain amount

accept the oﬀer

pay the amount

deliver the good

Figure 1: Message sequence chart of a story about a sale.

2.1 Signal Layer

The story describes four events, namely four acts per-

formed by two agents.

The ﬁrst two acts are easily

recognized as speech acts, but all these actions can be

teleologically interpreted as bringing some informa-

tional change into another agent. From this perspec-

tive, we can consider all of them as acts of commu-

nications, i.e. as messages going from a sender to a

receiver entity. Thus, the previous story can be illus-

trated using a communication diagram, for instance

as a message sequence chart (MSC).

Simplifying the

notation, we obtain the illustration in Fig. 1.

So far all seems easy, but the sale process de-

scribed before lacks some important details. For in-

stance, in a marketplace, paying and delivering are

physical actions. They produce some consequence

in the world: money, goods move from one place to

another. Furthermore, payment and delivery are ac-

knowledged by perception. Second, the buyer has ac-

cepted the offer, because he was somehow receptive

toward that kind of messages. Third, a buyer who al-

ready paid usually does not leave without taking the

good, just as the seller does not allow the buyer to go

away with an object without paying. In our story, all

goes well, so the narrative does not provide any el-

ement concerning these checks. However, this does

not mean necessarily that the buyer and seller had not

checked if everything was ﬁne. Fourth, sometimes a

buyer takes the good and then pays, sometimes the

order of actions is inversed. These four points reﬂect

characteristics which are left implicit in the story, and,

consequently, in the MSC:

• acts have side-effects on the environment (at the

very least, a transient in the medium transferring

the signal),

• an action consists of an emission (associated to an

There is also another implicit agent, the narrator, but

he will be neglected in this work, supposing he has a pure

descriptive intention.

MSCs are the basis of the sequence diagrams, one of

the behavioural diagrams used in UML. For further infor-

mation, see for instance (Harel and Thiagarajan, 2004).

pay

OUT

buyer seller

OUT

IN OUT

oﬀer

deliver

paid

delivered

world

Figure 2: Topology of the story.

agent) and of a reception (associated to a patient),

• certain actions have a closed-loop control: agents

perform some monitoring on expected outcomes,

• the sequence of events/acts in a story is often a

partial order, hidden by the linear order of the dis-

course.

Apart from the last point, which could be solved with

the UML “par” grouping for parallel constructs, we

have to ﬁnd alternative representations to help the

modeler in scoping and reﬁning the content of the

story.

2.1.1 Topology of the Story

Inspired by the Actor model (Hewitt et al., 1973),

we have drawn in Fig. 2 the topology of the story.

The topology serves as a still picture of the whole

case, and show how signals are distributed between

the characters. The little boxes are messages queues,

the lines identify communication channels. The story

describes which speciﬁc propositional content is used

in the exchanged messages. In order to take even-

tual side-effects into account, we introduced an ex-

plicit “world” actor, disjoining the emission from the

reception. The optional part of the communication is

visualized with dotted lines. The world would play

as intermediary entity also in case of broadcast mes-

sages.

2.1.2 Flow of the Story

Orthogonal to the topology, we deﬁne the ﬂow of the

story as the order in which events occurred. As a

ﬁrst deﬁnition, we may consider a story as a chain

of events (a strictly ordered set):

E = {e

, e

, ..., e

} (1)

In narratology this layer is usually called the fab-

ula: “a series of logically and chronologically re-

lated events [..]” (Bal, 1997). This name dates back

to Propp, which, altogether with the Russian formal-

ists, started considering each event in the the story as

With a similar spirit, communication acts performed

autonomously by the world actor can model natural events.

FromInter-agenttoIntra-agentRepresentations-MappingSocialScenariostoAgent-roleDescriptions

623

functional, i.e. a part of a whole sequence, necessary

to bring the narrated world from initial conditions to

a certain conclusion. Furthermore, speciﬁc circum-

stances may be described in correspondence to the oc-

currence of an event. As a result, a story corresponds

to the following chain:

→ C

→ ...

→ C

(2)

where e

are associated to transitions and C

is a set

of conditions assumed to continue at least until the

occurrence of e

Consequence and Consecutiveness. This deﬁni-

tion may look very simple, but the manifold relations

between consequence (logical, causal, ..) and consec-

utiveness (informed by time, ordering, ..) are actu-

ally very delicate to assess. Furthermore, two differ-

ent chronological coordinates coexist in a narrative: a

story-relative time, i.e. when the event has occurred

in the story, and a discourse-relative time, i.e. when

that event has been reported or observed.

In order to unravel this knot, we use a four steps

methodology to reconstruct the relations between the

elements of the story.

First, we elicit relevant abstractions which are

used in the interpretation. In particular, we deﬁne

an event/condition as free if the interpreter does not

acknowledge any relation

with another event or con-

dition in the story. We refer to such relations as de-

pendencies. Some dependencies are syntactic. For

instance, you can accept an offer only if there is an

offer, i.e. if an offer has been previously made. Oth-

ers are contextual to the domain. For instance, in a

web sale, payment usually occurs before delivery. In

all cases, dependencies can be used to put a strong

constraint on the ordering of events.

Second, there may be clues of the story-relative

time within the text. Time positions and durations

are usually meant to give some landmark to the lis-

tener. They are described in absolute or relative terms.

When a listener interprets them, it creates a relation

between events, contingent to the story. Such relative

positioning constitute the medium constraint.

Third, if we have no clues about dependencies,

or temporal relations between events, a possible se-

quence is at least suggested by the discourse-relative

time. This provides a weak constraint on the ordering

of free events.

If all three constraints are satisﬁed, we do not ex-

pect any concurrent events, at least within one story

Apart having occurred in the same story.

The story and discourse contingencies of the medium

and weak constraints become contextualities if they are en-

tailed by strong constraints.

s>b:oﬀer

(good, [for]amount)

b>s:accept

(oﬀer(good,

[for]amount))

s>b:deliver

(good)

b>s:pay

(amount)

s>w:deliver

(good, [to]s)

w>b: delivered

(good, [from]s)

b>w:pay

(amount,

[to]s)

w>s:paid

(amount,

[from]b)

Figure 3: Flow of the story.

frame.

However, it is easy to object to such a strict

determination.

Consequently, at the fourth step, we weaken the

previous strict temporal constraints (e.g. from e

i+1

to e

i+1

≥ e

) in two cases: (a) dependencies can be

associated to no-time-consuming processes (e.g. log-

ical equivalences); (b) events may occur simultane-

ously, when triggered by parallel sub-systems. Fur-

thermore, the medium and weak constraints refers to

contingent relations (according to the modeler). In or-

der to be able to compare the internal structure of sto-

ries, we can neglect them. With these modiﬁcations,

the set E deﬁned in (1) is a partially ordered set.

Let us take our case story. There is a relation of

syntactic necessity between offer, acceptance and per-

formance. In addition, there are two agents. These en-

tities can be considered as parallel systems, that may

concurrently interact with the world. Therefore, with-

out further contextual speciﬁcation, payment and de-

livery are concurrent events.

Visualization. A simple way to visualize the ﬂow

of the story is by the use Petri nets, as we did in

Fig. 3. We opted for a practical naming of places:

sender>receiver:content. At this point, places

represent messages, associated to speech acts. Ac-

tions are like “compacted” into transitions.

The main scope of the ﬂow is to preserve the story

synchronization. Further layers may be integrated,

increasing the granularity of the description, main-

taining the previous points of synchronization, in the

same spirit of hierarchical Petri nets (Fehling, 1993).

For instance, in Fig. 3, we have disjoint the generation

of the message from its reception using an intermedi-

ate “world” actor, as we did in the topology.

A more complex story may consist of many frames.

For “frame”, we consider a sub-story that follows the Aris-

totelian canon of unities of time, space and action.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

624

seller

buyer

!sell(good, [for]amount)

!buy(good)

oﬀer(good, [for]amount)

?acceptable(oﬀer)

critical

[acceptable(oﬀer)]

accept(oﬀer)

critical

[has(buyer, amount)]

pay(amount)

critical

[has(seller, good)]

deliver(good)

bought(good)

sold(good, [for]amount)

Figure 4: MSC with intentions and critical conditions.

3 AGENTIC

CHARACTERIZATION

In the previous section we referred to the messages

exchanged in a social scenario, and narrated through a

story. When we interpret such messages, however, we

apply an intentional stance just as we do in our experi-

ential life: we read the actors as intentional agents, at-

tributing them beliefs, desires and intentions. This ex-

post intentional interpretation of the story results in a

decomposition of the plans followed by the agents. In

addition, in order to trigger or enable the performance

of the reported acts, there may be other relevant con-

ditions or hidden acts to be taken into account. They

could have been left implicit by the narrator, but plau-

sibly they are known to the target audience of that

narrative. This is the main assumption on which our

work is based: any reader/modeler can always pro-

vide one or several reconstructions of what happens

behind the signal layer, supported by common knowl-

edge and domain experts (or, at least, interviewing the

narrator himself, if needed).

3.1 Acquisition Methodology

As we did before, we start the acquisition using a

MSC diagram. A possible outcome of the interpre-

tation is in Fig. 4, where we introduce adequate ex-

tensions.

First, we consider externalized intents (with a “!”

preﬁx) as the events triggering the processes of buy-

ing/selling. The ﬁnal outcomes of those actions are

then reported with output messages as well, at the end

of the chart. Second, we add eventual hidden acts. In

our case, we know that a buyer usually accepts an of-

fer only after positively evaluating it.

Third, we use

the critical grouping to highlight which conditions (in

addition to sequential constraints) are necessary for

the production of that message.

To sum up, in our story, we add that: (a) the buyer

performs an evaluation of the offer (evaluation ac-

tion), (b) the buyer accepts the offer if it is acceptable

for him (acceptability condition), (c) the buyer pays

(the seller delivers) if he owns the requested money

(good) (ownership condition).

The MSC diagram in Fig. 4 furnishes a good sum-

mary of the story: the inputs/outputs provide an inten-

tional characterization, the vertical bars indicate the

ongoing activities, while the messages refer to suc-

cessful acts of emission and reception, whose occur-

rence is constrained by the critical conditions. Unfor-

tunately, further reﬁnement is necessary to cover the

basic ﬁgures encountered in an operational setting.

There is no separation between emission and recep-

tion, and between epistemic and ontological. For in-

stance, we cannot distinguish between a case in which

the buyer thinks he has not enough money, and an-

other in which he thinks that there is enough on his

bank account, but the bank does not “agree” with him.

A simple solution to this problem would be to add

intermediate actors (e.g. the bank), localizing where

the failure occurred. In a complex case, however, the

resulting visualization may be overloaded. In the fol-

lowing sections we will therefore introduce some pat-

terns to be attached to the ﬂow of the story. Instead

of using just one visualization, our approach aims to

provide alternative representational cuts.

3.1.1 Hierarchical Tasks

First, we elicit the hierarchical decompositions of ac-

tivities performed by the agents. These serve as basic

schemes for the behavioural characterization of the

agents, and use hierarchical, serial/parallel constructs.

In practice, this is obtained by identifying, in the story

ﬂow, the activities of the actor as agent (emitter) and

patient (receiver), and relating them according to their

dependencies. Fig. 5 reports the result of this step for

the buyer.

3.1.2 Emission and Reception

Second, activities are anchored to messages. This is

The absence of evaluation is symptomatic of a combine

scheme: the buyer performs mechanically the acceptance

in order to advance the interests of another element of his

social network.

FromInter-agenttoIntra-agentRepresentations-MappingSocialScenariostoAgent-roleDescriptions

625

buying

listening(oﬀer)

start

action

layer

accepting

listening(delivered)

paying

end

action

layer

[impulse]

evaluating

Figure 5: Hierarchical decomposition of tasks.

E C

start

action

layer s

start

message

layer

start

action

layer b

b:listening(content)

s:telling(content)

s>b:content

s>w:content

w>b:content

end

message

layer

end

action

layer b

end

action

layer s

generation

synchronization

perception

Figure 6: Communication pattern.

a delicate phase: we want to maintain the synchro-

nization given by the story and the dependencies as-

sociated to the activities. Fig. 6 reports our solution

(applied on a single message) which explicitly divide

emission from reception. The proposed Petri net is

complete and well-formed. It is scalable to multiple

agents, adding a reception cluster (e.g. w>b:content,

perception, etc.), in order to connect the message to

each agent that is reachable by the communication.

Note: partial orderings may hold independently in

the story ﬂow and in the activity diagrams. In this

case, for instance, we do not know a priori if payment

occurs before delivery (as acts), as we do not know

if the buyer pays before monitoring the delivery (as

actions), or vice-versa, or simultaneously.

3.1.3 Illocutionary Acts

Third, we recognize the practical effects of messages.

Beside of being signal (or a locutionary act), each

message is associated to an illocutionary act, and

then, when it is put in a computational form, it should

integrate some pragmatic meaning.

For simplicity, we consider only four types of per-

formatives: assertions, commissives, commands and

inquiries. Moreover, we interpret commitments as

obligations to the self and commands as attempts to

The simultaneity of actions is evident when consider-

ing a collective agency. However, considering an individual

physical agent, we can assume at least that, concurrently to

the performance of a deﬁnite activity, he maintains some

listening activities as well.

instill obligations into the receiver. Considering our

story, we have:

• the offer is a conditional promise: the seller com-

mits to deliver the good to a buyer who commits

to pay his price;

• the acceptance is a promise: the buyer commits to

pay a given amount;

• the payment can be interpreted as a command per-

formed by the buyer to the world, plus an asser-

tion performed by the world to the seller;

• the delivery can be modelled similarly to the pay-

ment.

3.1.4 Action and Power

Fourth, activities are used as anchors for cognitive,

motivational and institutional elements, informed by

the illocutionary content of the messages. Our intu-

ition is that obligations are prototypical motive for

actions, while power representations are handled at

epistemic level. Therefore, in our representation, we

refer in total to four layers, each of which addresses

speciﬁc components:

• the signal layer — acts, side-effects and failures

(e.g. time-out): outcomes of actions,

• the action layer — actions (or activities): perfor-

mances intended to bring about a certain result

(the action layer),

• the intentional layer — intentions: commitments

to actions, or to nested intentions,

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

626

[impulse]

pay

end

message

layer

end

agentic

layer

paid

contextualization

start

agentic

layer

start

message

layer

obligation to pay

intention to pay

paying

affordance

dispositionmotivationmotivation

motive

b actually has

the power to pay

(e.g. he actually

owns the money)

b thinks he has

the power to pay

(e.g. he thinks he

owns the money)

b intends

to be compliant

b knows how to infer

the pragmatic meaning

of an acceptance

acceptance

Figure 7: Full action pattern associated to payment.

• the motivational layer — motives: events trigger-

ing the creation of intentions.

The last three layers compose the agentic layer. The

closure of the sensing-acting cycle of the agent is

guaranteed by the fact that certain signals, when per-

ceived by agents, becomes motives for action. But

how the motive is translated in intention(s)? How an

intention is transformed in action(s)? What permits

that an action effectively produces a certain result?

These questions can be answered introducing ad-

ditional elements, dual to the previous components,

such as:

• dispositions: contextual alignments of the agent

with the environment (consisting of other agents

and of the world actor) in respect to the actions he

performs,

• affordances: perceived alignments of the agent

with the environment in respect to his intents,

• motivations: mental states catalyzing the creation

of intentions.

Affordance and Disposition. Taking an intentional

stance, all behaviours become intent-oriented. If

the agent thinks that the environment affords his be-

haviour, he also thinks he has the power to achieve

the goal associated to that behaviour.

Therefore, af-

fordance practically corresponds to perceived power.

In the literature—see for instance (Chemero, 2003)—

the term affordance often refers both to the perceived re-

lation animal/environment and to the effective (or disposi-

tional) relation. In this paper we will associate the term

affordance only to the ﬁrst meaning, preferring the term dis-

position for the second.

This assertion has no relation with consciousness. In-

tent can be implemented by design. A certain tool, or even

a business process, is designed to achieve a certain goal in

Similarly, disposition is connected to actual power:

it is a precondition to the consequences that a certain

action of the agent will imply.

In the light of this analysis, we observe how these

categories corresponds to the critical conditions re-

ported on the MSC (e.g. acceptability, ownership in

Fig. 4). The subjective evaluation of each of these

conditions gives to the agent the affordance of the ac-

tion resulting in that message. However, if the affor-

dance is a sufﬁcient condition for the performance of

the action, it is only necessary for the intended out-

come, where the contextual disposition plays a role.

Considering our story, if the buyer starts buying, (it is

like) he is assuming he has the power to do it—said

equivalently, according to him the affordance of buy-

ing holds. Despite of his intent, however, the seller

may be a fake seller, aiming to get the money without

delivering. In this case, the disposition for a success-

ful transaction would not hold, blocking the normal

completion of the sale.

Motive and Motivation. In our framework, moti-

vation refers to some mental condition that makes the

agent sensitive to a certain fact, which becomes the

motive for starting an action. As we observed be-

fore, obligations are prototypical reason for actions.

Despite of that, not all obligations are followed by a

performance. People comply with obligations when

they have some motivation to: it may be for habit,

convenience, respect to authority, or for fear of rein-

forcement actions. Motivations however often remain

implicit in the story.

a certain environment, but artefacts or processes in them-

selves are not aware of such goal or environment. Intent

and affordance may transcend actual performance, but they

still exist.

FromInter-agenttoIntra-agentRepresentations-MappingSocialScenariostoAgent-roleDescriptions

627

Visualization. To sum up the concepts introduced

so far, we have reported in the Petri net in Fig. 7 a pos-

sible reconstruction of the step of payment performed

by the buyer. To reduce the visual burden, we have

combined the layers action, intentional, motivational

in one agentic layer. Nevertheless, it would be suf-

ﬁcient to add starting and ﬁnishing synchronization

places for each layer to have a complete multi-layered

model. The picture maintains in fact the vertical or-

ganization of our conceptualization: it is easy to rec-

ognize to which layer each element belongs.

The triggering motive is the acknowledgement of

an acceptance (this could be the buyer’s own accep-

tance to an offer, or the reception of a seller’s ac-

ceptance to his own offer). The illocutionary con-

tent of the offer/acceptance entails the duty to pay of

the buyer. This duty to pay is followed for instance

if the buyer desires to be compliant. An obligation

is then formed and used to construct the correspon-

dent intention. The intention, if there is an action or a

course of actions (i.e. plan) which is afforded by the

environment (e.g. the buyer thinks he owns enough

money), supports the selected performance. Finally,

if the action is performed in a correct alignment (e.g.

the buyer owns enough money), it results in the ex-

pected act.

In this type of Petri net we observe an interest-

ing pattern — see Fig. 7: certain transitions are con-

nected both to impulse and to persistent places. The

ﬁrst identify events (the occurrence of change), and

the second conditions (the existence of a continuity).

3.2 Model Validity

Each observed scenario can be explained by several

interpretations: as the reader probably thought while

reading our examples, there is not only one way of

reconstructing the mechanisms that bring about the

production of messages. The story provides only a

foreground, which the interpreter can anchor to al-

ternative backgrounds, explaining and completing it

further. Further development of this research will be

the study of a subsumption relation between scenarios

(analysing their Petri nets) and consequently a dis-

tance operator which considers partial overlaps be-

tween them.

In this work, however, our focus is not the “sci-

entiﬁc” validity of the systems integrated in the back-

ground, i.e. how far they represent what is the case

in the experiential world; we accept that the choice of

backgrounds is informed by the ontological commit-

ments of the interpreter (e.g. judges giving alternative

interpretations to a legal case). In our perspective, if

an interpretation produces all the messages provided

in the foreground, then, such model is valid. The pro-

cess of selection becomes a matter of justiﬁcation.

4 INTRA-AGENT SYNTHESIS

In the previous section, we have augmented the de-

scription of inter-agent interactions with intentional

(and conversational) concepts. In this section, we will

provide elements of the integration of our framework

with current practices in MAS.

4.1 Agent-roles

Rationality is commonly deﬁned as the ability of the

agent to construct plans of actions to reach a goal,

possibly referring to a hierarchical decomposition of

tasks. In our case, agents do not deliberate, or bet-

ter, the decisions they deliberate have been already

taken.

Consequently, their behaviour is fully deter-

ministic. Therefore, instead of considering a full ac-

count of agency, we opted to base our framework on a

more constrained concept: the agent-role, which inte-

grates the concepts of narrative role and institutional

role in a intentional entity.

Agent-roles are self-other representations (Boer

and van Engers, 2011a), i.e. used to interpret, plan

or predict oneself’s or others’ behaviour. They are

indexed by: (a) a set of abilities, (b) a set of suscepti-

bilities to actions of others. In a social scenario, both

abilities and susceptibilities become manifest as mes-

sages exchanged with agents playing complementary

agent-roles. Thus, the topology of a scenario (e.g.

Fig. 2) provides a fast identiﬁcation of these indexes.

Differently from objects (and actors), however, agent-

roles are entities associated also to motivational and

cognitive elements like desires, intents, plans, and be-

liefs.

4.2 From Scenarios to Agents

A scenario agent is an intentional agent embodying

an agent-role. It is deterministic in its behaviour: all

variables are either set, or determined by messages

from other agents. This determinism, consequent to

the internal description of the agent, constrains the

temporal order of messages between agents, but not

the behaviour of an entire multi-agent system. In sec-

tions 2 and 3 we have constructed and used a message

layer in order preserve the synchronization given by

Part of the results of the deliberation are reported in the

story, possibly together with elements of the deliberation

process. Despite of these traces, however, the task decom-

position occurs mostly in the mind of the modeler.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

628

the case. If we remove this layer, the system becomes

non-deterministic. For instance, in a case in which

there are two buyers and a seller, we do not know a

priori who is the ﬁrst buyer to accept.

Agents may communicate only with agents in

their operating range, i.e. whose message boxes are

reachable to them. Each agent may have different

message boxes, and each of them can be epistemically

associated to an identity.

4.3 Computational Implementation

For its structural determinism, the behaviour of an

agent-role can be described in terms of rules, translat-

ing logical and causal dependencies. As we already

observed analyzing Petri nets like in Fig. 7, some tran-

sitions are connected to places associated to impulse

events and to places named after continuing condi-

tions. This conﬁguration easily relates our represen-

tation to event-condition-action (ECA) rules, com-

monly used in reactive systems. In respect to MAS

theory, this connection has been exploited already

in AgentSpeak(L) (Rao, 1996), a logic programming

language for cognitive agents, extended and opera-

tionally implemented in the platform Jason (Bordini

et al., 2007). The connection of AgentSpeak(L) with

Petri nets has been extensively studied in (Behrens

and Dix, 2008), with the purpose of performing MAS

model checking using Petri nets. In the present work,

however, we have a completely different objective:

we started from a representation of the scenario on a

MSC chart, we reﬁned it with Petri nets patterns, and

now we want to extract from this representation the

correspondent agent-role descriptions (as agent pro-

grams).

4.3.1 Overview of AgentSpeak(L)/Jason

In order to proceed, this section presents very brieﬂy

two important constructs of AgentSpeak(L)/Jason.

The ﬁrst is the ECA rule associated to the activation

of a goal. Put in words, with an imperative ﬂavour:

in order to reach the goal, if certain conditions are

satisﬁed, perform this plan of actions. The equivalent

code is something like:

+! goal : cond i ti o nA & ... & con d it i onZ

<- actio n A ; . . . ; act i onZ .

Conditions represent what the agent thinks should be

true, in that very moment, in order to be successful in

executing his plan. The propositional content is writ-

ten in a Prolog-like form. In addition, Jason permits

Current MAS platforms refer instead to ontological

identities.

the attachment of annotations (a sort of optional pred-

icated parameters, expressed within squared brack-

ets). As for actions, they are either direct operations

with the environment, or !g (triggering the activation

of a goal g), +b, -b (respectively adding and removing

the belief b from the belief base).

The second construct we present is an ECA rule

concerning the addition of a belief: when a certain

belief is added, if certain conditions are satisﬁed, per-

form this plan of actions.

+ b e l ief : con d it i onA & ... & co n di t ion Z

<- actio n A ; . . . ; act i onZ .

Although similar, these rules have different se-

mantics. The motivational component !goal disap-

pears when the plan ends successfully and also when

it fails (in this case, the event -!goal is triggered).

The knowledge component belief is instead main-

tained.

4.3.2 Case Example: Buyer’s Payment

As example of synthesis, we are now able to trans-

late the interpretation of payment illustrated in Fig. 7.

This is an excerpt of the code of the buyer agent-

role

+! acc e p t ( o f fer ( Good , Am o unt )

[ s o u rce ( Se ller )])

<- . se n d ( Seller , tell ,

acce p t ( offer ( Good , Amount )));

+ obl ( p a y _to ( Amount , Sel l er )).

+ obl ( p a y _to ( Amount , Agent )) < -

! p a y _to ( A mount , A g e n t );

- o bl ( pay_to ( Amount , Agen t )).

+! pay _ t o ( Amount , Agent )

: ownin g ( Money ) & M o n e y >= Amoun t

<- . se n d (w , achieve ,

pay_ t o ( Amo unt , Agent ));

+ p a id_t o ( Am ount , Agent ).

For completeness, we included in the ﬁrst rule also

the generation of the speech act of acceptance. Ne-

glecting this action, we have three rules, hierarchi-

cally dependent, and the last one performs the speech

act for the payment. It is easy to observe a strong cor-

respondence between these elements and the layers in

Fig. 7. The ﬁrst rule acknowledges the acceptance and

generates the event/condition concerning the obliga-

tion (motivational layer), the second rule transforms

.send/3 is an action provided by the MAS platform

that generates speech acts. The ﬁrst parameter is the target

agent, the second is the illocutionary force (tell for asser-

tions, achieve for directives), the third is the propositional

content.

FromInter-agenttoIntra-agentRepresentations-MappingSocialScenariostoAgent-roleDescriptions

629

the obligation of paying in intention of paying (moti-

vational layer); the third rule checks the affordance

related to the intent (intentional layer) and, if this

evaluation is positive, performs the paying action (ac-

tion layer). The action is then externalized to a com-

munication module of the agent, interacting with the

world/environment, which in turn will generate the

actual consequence (signal layer).

To conclude, we observe that in this description

the belief of having paid can be only partially aligned

with the ontological reality. From the perspective of

the agent, if the action has not failed, it is natural

to think that it has been successful. In reality, how-

ever, something may block the correct transmission

of the act to its beneﬁciary (e.g. a failure in the bank

databases). This extent of such alignment is related

to the focus of the feedback process checking the per-

formance.

5 DISCUSSION

The modeling exercise running through the paper

served as an example of operational application of a

knowledge acquisition methodology targeting socio-

institutional scenarios. Each representation we con-

sidered (MSC, topology, Petri net, AgentSpeak(L)

code) has shown its weakness and strengths in this re-

spect. Furthermore, the cross-relations between them

are not simple isomorphisms. Despite of these difﬁ-

culties, we think that using alternative visualizations

is a way to achieve a more efﬁcient elicitation (target-

ing also non-IT experts). In this line of thought, we

plan to implement and assess an integrated environ-

ment for knowledge acquisition; the scalability of the

methodology should be supported by the introduction

of an adequate subsumption relation between stories,

allowing faster elicitation of models.

From a higher-level perspective, the present work

connects scenario-based (or case-based) modeling

with multi-agent systems technologies. The idea at

the base is that, in order to acquire representations of

social behaviours, we need cases to be valid models,

and we can validate them by their execution.

(Mueller, 2003) observes that, although several

story understanding programs—starting from BORIS

(Charniak, 1972)—have used sort of multi-agent sys-

tems for their internal representation, this choice is

not easy for the programmer: such agents are difﬁcult

to write, maintain, and extend, because of the many

potential interactions. His experience matched with

ours. However, we think that the connection of agent-

based modeling with MAS is too strong and impor-

tant to be easily discarded. As longer-term objective,

we aim to couple on the same simulation framework

designed systems (e.g. IT infrastructures) and repre-

sentations of known social behaviours.

Scenario-based Modeling. MSCs (and collections

of them, e.g. HMSCs) were standardized as support

for the speciﬁcation of telecommunication software,

in order to capture system requirements and to collect

them in meaningful wholes (Harel and Thiagarajan,

2004). Later on, other extensions, like LSCs (Damm

and Harel, 2001) and CTPs (Roychoudhury and Thi-

agarajan, 2003), were introduced to support the auto-

matic creation of executable speciﬁcations. The basic

idea consists in collecting multiple inter-object inter-

actions and synthesizing them in intra-object imple-

mentations. In principle, we share part of their ap-

proach. Our work promotes the idea of using MSCs,

although integrated with intentional concepts. How-

ever, in their case, the target is a speciﬁc closed sys-

tem (to be implemented), while in our case, a sce-

nario describes an existing behavioural component of

an open social system. At this point, we are satisﬁed

by transforming the MSC of a single case in the corre-

spondent agent-roles descriptions. The superposition

of scenarios, with the purpose of associating them into

the same agent-role, is an open research question.

Story Understanding. AI started investigating sto-

ries in the ’70s, with the works of (Charniak, 1972),

(Abelson and Schank, 1977), introducing concepts

like scripts, conceptual dependency (Lytinen, 1992),

plot units (Lehnert, 1981). The interest towards this

subject diminished in the early ’80s, leaking into other

domains. (Mateas and Sengers, 1999) and others tried

a refocus in the end of ’90s, introducing the idea of

Narrative Intelligence, but again, the main stream of

AI research changed, apart from the works of Mueller

(e.g. (Mueller, 2003)). All these authors, however,

are mostly interested in story understanding. We are

investigating instead the steps of construction of what

they called script (Abelson and Schank, 1977). Ac-

cording to our perspective, common-sense is not con-

structed once, in a script-like knowledge, but emerges

as a repeated pattern from several representations.

Furthermore, we explicitly aim to take account of the

integration of fault and non-compliant behaviours, in-

creasing the “depth of ﬁeld” of the representation.

Computational Implementation. Reproducing a

system of interacting subsystems needs concurrency.

Models of concurrent computation, like the Actor

model (Hewitt et al., 1973), are implemented today in

many development platforms. In our story-world, this

solution would be perfect for objects. We would need

instead to add intentional and institutional elements in

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

630

order to implement agent-roles. The connection with

another programming paradigm (intended to handle

concurrency) will plausibly solve most of the prob-

lems of scalability that usually haunt MAS platforms,

often developed by a logic-oriented community.

Relevance. The Agile methodology for public ad-

ministrations (Boer and van Engers, 2011a) intro-

duces the concept of agent-role, and targets the ex-

ploitation of agent-role descriptions, as components

of a knowledge-base corresponding to the deep model

(Chandrasekaran et al., 1989) of a social reality. Such

a model can be used to feed design and diagnostic

(Boer and van Engers, 2011b) applications with the

purpose of supporting the activity of organizations.

The legal system in many areas presupposes the use

of informal or semi-formalized models of human be-

haviour in order to operate. If we aim to support an

administrative organization on those points, ABM is

the most natural choice. In doing this, we recognize

we are going in opposition to the current drift from

agent-based modeling to computational social sci-

ence (Conte and Paolucci, 2011). However, because

of the generative aspect of the agent-role concept, our

contribution is relevant to research in behaviour ori-

ented design (Bryson, 2003) or similar, usually ap-

plied to robotics and AI in games.

REFERENCES

Abelson, R. P. and Schank, R. C. (1977). Scripts, plans,

goals and understanding. Lawrence Erlbaum Ass.

Bal, M. (1997). Narratology: Introduction to the Theory of

Narrative. University of Toronto Press, 2nd edition.

Boer, A. and van Engers, T. (2011a). An agent-based legal

knowledge acquisition methodology for agile public

administration. In Proc. of 13th Int. Conf. ICAIL 2011,

171–180.

Boer, A. and van Engers, T. (2011b). Diagnosis of Multi-

Agent Systems and Its Application to Public Admin-

istration. In Business Information Systems Workshops,

Lecture Notes in Business Information Processing,

97:258–269.

Behrens, T. M. and Dix, J. (2008). Model checking multi-

agent systems with logic based Petri nets. Annals of

Mathematics and Artiﬁcial Intelligence, 51(2-4):81–

121.

Bordini, R. H., H

ubner, J. F., and Wooldridge, M. (2007).

Programming multi-agent systems in AgentSpeak us-

ing Jason. John Wiley & Sons Ltd.

Preliminary experiments of translating Jason on a func-

tional programming paradigm, like for instance in Erlang

(Diaz et al., 2012), have proven to be computationally very

efﬁcient.

Bryson, J. (2003). The behavior-oriented design of modu-

lar agent intelligence. Agent technologies, infrastruc-

tures, tools, and applications for e-services, 61–76.

Chandrasekaran, B., Smith, J. W., and Sticklen, J. (1989).

’Deep’ models and their relation to diagnosis. Artiﬁ-

cial Intelligence in Medicine.

Charniak, E. (1972). Toward a model of children’s story

comprehension. Technical report, MIT.

Chemero, A. (2003). An Outline of a Theory of Affor-

dances. Ecological Psychology, 15(2):181–195.

Conte, R. and Paolucci, M. (2011). On Agent Based Mod-

elling and Computational Social Science. Available at

SSRN 1876517.

Damm, W. and Harel, D. (2001). LSCs: Breathing life into

message sequence charts. In Formal Methods in Sys-

tem Design, 45–80.

Diaz, A. F., Earle, C. B., and Fredlund, L.-A. k. (2012).

eJason: an implementation of Jason in Erlang. Proc.

of 10th Workshop PROMAS 2012.

Fehling, R. (1993). A concept of hierarchical Petri nets with

building blocks. Advances in Petri Nets, 674:148–168.

Harel, D. and Thiagarajan, P. S. (2004). Message sequence

charts. UML for Real: Design of Embedded Real-time

Systems, 75–105.

Hewitt, C., Bishop, P., and Steiger, R. (1973). A univer-

sal modular actor formalism for artiﬁcial intelligence.

Proc. of IJCAI 1973, 235–245.

Lehnert, W. G. (1981). Plot units and narrative summariza-

tion. Cognitive Science, 5(4):293–331.

Lytinen, S. L. (1992). Conceptual dependency and its de-

scendants. Computers & Mathematics with Applica-

tions, 23(2):51–73.

Mateas, M. and Sengers, P. (1999). Narrative intelligence.

In AAAI Fall Symposium on Narrative Intelligence.

Mueller, E. T. (2003). Story understanding through multi-

representation model construction. In Text Meaning:

Proc. of HLT-NAACL 2003 Workshop, 46–53.

Rao, A. S. (1996). AgentSpeak (L): BDI agents speak out in

a logical computable language. In Proc. of 7th Euro-

pean Workshop on Modelling Autonomous Agents in

a Multi-Agent World.

Roychoudhury, A. and Thiagarajan, P. (2003). Communi-

cating transaction processes. Proc. of 3rd Int. Conf. on

Application of Concurrency to System Design, 157–

166.

Sileno, G., Boer, A., and Van Engers, T. (2012). Analysis

of legal narratives: a conceptual framework. In Proc.

of 25th Int. Conf. JURIX 2012, 143–146.

FromInter-agenttoIntra-agentRepresentations-MappingSocialScenariostoAgent-roleDescriptions

631