MODEL OF KNOWLEDGE SPREADING FOR MULTI-AGENT

SYSTEMS

D. Oviedo, M. C. Romero-Ternero, M. D. Hern´andez, A. Carrasco, F. Sivianes and J. I. Escudero

Departamento Tecnolog´ıa Electr´onica, Universidad de Sevilla, Sevilla, Spain

Keywords:

Multi-agent system, Knowledge sharing, Knowledge spreading, Ontologies (artiﬁcial intelligence), Open sys-

tems, Software engineering, Agent communication, Automatic control system, Interoperability model.

Abstract:

This paper presents a model to spread knowledge in multiagent-based control systems, where simplicity, scal-

ability, ﬂexibility and optimization of communications system are the main goals. This model not only implies

some guidelines on how the communication among different agents in the system is carried out, but also de-

ﬁnes the organization of the elements of the system. The proposed model is applied to a control system of

a solar power plant, obtaining an architecture which optimizes agents for the problem. The agents in this

system can cooperate and coordinate to achieve a global goal, encapsulate the hardware interfaces and make

the control system easily adapt to different requirements through conﬁguration. The model also includes an

algorithm that adds new variables in the communication among agents and enables ﬂow control knowledge in

the system.

1 INTRODUCTION

A goal of control systems is to enable the integration

of different types of devices in a scalable and ﬂexible

system. The problem is how communications among

the different parts of the system is organised and op-

timized (Huget, 2003).

The control systems have evolved to a complex

system that employs more and more equipment and

sensors. The hardware diversity of the sensors and

actuators greatly affects the portability of the con-

trol system and the complexity may differ from each

other. A solar powerplant control system (CARISMA

Project) is inserted in this paper. It has a distributed

input-output agent architecture to accomodate the

changing requirements. This distributed intelligent

agent architecture provides ﬂexible and scalable ways

to integrate the different sensors and actuators. The

design goal is to create a system architecture that

is general enough to support many different kinds

of sensors and actuators, while being distributed and

scalable.

To resolve problems in the control system it is nec-

essary have knowledge and experience working in the

ﬁeld. In the case presented, the solutions to control

problems will be the responsibility of different agents

in the system. The agents should handle the problem

domain knowledge and be able to communicate this

knowledge in order to provide efﬁcient solutions and

recommendations. In addition to the system architec-

ture, we propose a model for spreading the knowl-

edge of agents within this architecture. This model

presents an organizational scheme of agents and a

global communication protocol. The model was con-

structed to support a wide range of control systems;

however, this paper focuses on its application in the

control of a solar power plant.

2 PROBLEM DOMAIN

The problem discussed here belongs to the ﬁeld of

distributed control systems, particularly in the prob-

lem of knowledge spreading in distributed network

agents of control. An introduction to the subject and

issues can be found at (Li et al., 2008), (Yang et al.,

2009). Control systems based on the theory of mul-

tiagents, and restricted to speciﬁc domains have been

developed. For instance, one of the early works by

(Wang Junpu, 2000) discusses the feasibility of agent-

based distributed hierarchical intelligent control. An-

other study examines the modeling of multiagent con-

trol for energy infrastructures (Sebastian Beer, 2009)

that presents an agent-based control system for dis-

tributed energy resources in low voltage power grids.

Knowledge representation mechanisms primarily

326

Oviedo D., C. Romero-Ternero M., D. Hernández M., Carrasco A., Sivianes F. and I. Escudero J. (2010).

MODEL OF KNOWLEDGE SPREADING FOR MULTI-AGENT SYSTEMS.

In Proceedings of the 12th International Conference on Enterprise Information Systems - Artiﬁcial Intelligence and Decision Support Systems, pages

326-331

DOI: 10.5220/0002973803260331

 SciTePress

used for sharing of knowledge among agents, has

also been a focus of research. For instance, the use

of mathematical logic, for knowledge representations

and exchange among agents has been explored in

(Alessio Lomuscio, 2000). However, our focus is on

using ontologies for knowledge.

Within this domain, the most important aspect

that we discuss concerns the spread of knowledge

in control systems based on multiagents. In (Mara

Adela Grando, 2006) the study of knowledge trans-

fer and action restriction among agents in multiagent

systems founded on the deﬁnition of patterns of di-

alogues between groups of agents, are expressed as

protocols. In this paper we focus in the organization

of the elements of the system and the ﬂow of infor-

mation within the system, with the aim of creating a

simple and optimized model.

3 MODEL OF KNOWLEDGE

SPREADING

In multi-agent systems, one important aspect is the

sharing of knowledge among agents. To share knowl-

edge it is necessary to conceptualize the problem

domain and this should be common to all agents

(Knowledge Representation). It is also necessary to

deﬁne the processes for sharing knowledge or acquir-

ing knowledge by an agent. The following section

presents our model for the implementation of these

aspects discussed.

3.1 Knowledge Representation in the

Agents

Traditionally, agent knowledge representation is done

through what is called Ontologies (Colomb, 2007).

Ontologies deﬁne a set of elements as Predicates, Ac-

tions of agents, Concepts, Primitives, Expressions,

Variables, etc. In the case of control systems, the em-

phasis is mainly on the concepts and actions that an

agent could handle while other elements are more in-

tended for communication tasks.

The concepts can be easily represented by data

structures. In our CARISMA project, these concepts

will refer to the different variables that are neces-

sary to manage in order to make a solar panel con-

trol. These concepts can be stored locally on each

agent or centralized, but in any case, the concepts

must be shared by agents, so that they can commu-

nicate. The mechanisms used to update concepts can

be based on simple broadcasting of messages, updat-

ing a central database accessed by the various agents

to learn the concepts used or in distribute mechanisms

(Sebestyenova, 2005).

The actions of the agents are modeled by expert

systems (Yanping Du, 2005), using facts and rules.

These rules represent real actions on devices such as

reporting recommendations, faults or alarms in the

system that have been located by one or more agents.

Also these rules in an agent may represent the action

required to share knowledge with other agents or to

update their own knowledge. The mechanisms for up-

dating expert systems agents are outlined in the next

sections.

3.2 Knowledge Sharing by Agents

To spread the knowledge in a multi-agent system, it is

necessary that agents are equipped with mechanisms

for communication of this knowledge through the net-

work that interconnects them. These mechanisms are

usually based on communication primitives (Walton,

2003; Milner, 1994) deﬁned in the ontology of the

system.

In order to improve the reliability of the recom-

mendations and proceedings of a multi-agent con-

trol system, we propose to label the knowledge that

spreads through the network. Such labeling would

be based on two concepts: the reliability of infor-

mation and the reputation of the agent that has com-

municated the information. The reliability indicates

the level of credibility of the received information by

an agent, while the reputation indicates the degree of

trust that the recipient agent has regarding the agent

which transmitted the information. These labels can

be represented as a percentage. In the transmission

of information, the transmitter adds informationabout

the reliability to the knowledge and labels the reputa-

tion for the receiver.

3.3 Knowledge Acquisition by Agents

When an agent receives a message from another agent

which carries new knowledge or an update existing

knowledge, there must have a process / task, designed

to update the expert system of the destination agent

(Dongliang et al., 2008). These processes operate di-

rectly on the agent’s knowledge base, so additionally

we propose an algorithm that considers the labels that

are associated with knowledge: reliability and reputa-

tion.

In the algorithm the variable newK represents the

new knowledge, SA is the source agent and DA is the

destination agent, newFb and fb are the new ﬁabil-

ity and the ﬁability label respectively, rp is the rep-

utation label of source agent .OldK refers to the old

MODEL OF KNOWLEDGE SPREADING FOR MULTI-AGENT SYSTEMS

327

knowledge stored in the destination agent. Depend-

ing on the results of that algorithm, it is determined if

the knowledge update process will take place or not

in the agent.If the algorithm opens the way to update

the information for the agent, it also indicates the la-

bels of reliability and reputation of the new knowl-

edge gained.

Below is the algorithm written in pseudocode and

details of its operation , as well as the set of functions

that are used in it.

Algorithm 1. acquisitionKnowledge.

if not hasKnowledge(DA,newK)

if acceptByReputation(DA,SA)

newFb=newK.fb/SA.rp

insertKnowledge(DA,newK,newFb)

else

discardKnowledge(DA,newK)

endif

else

if acceptByReputation(DA,SA)

if(oldK.fb <= newK.fb)

newFb=newK.fb/SA.rp

updateKnowledge(DA,oldK,newK,newFb)

else

if confirmKnowledge(DA,newK)

newFb=newK.fb/SA.rp

updateKnowledge(DA,newK,newFb)

else

discardKnowledge(DA,newK)

endif

else

discardKnowledge(DA,newK)

endif

If the information is new knowledge (not has-

Knowledge) then the destination agent will check the

reputation it has of the source agent (acceptByRep-

utation, for example based on a threshold). When

the new knowledge is accepted by reputation, then

the algorithm gets the value of the new label of re-

liability (such as the division between the reliabil-

ity of information and the value of reputation of the

source agent) and the new knowledgeis inserted in the

agent’s knowledge base stations (insertKnowledge).

In other cases, new knowledge is discarded (discard-

Knowledge). If the information is an update of knowl-

edge that is already possessed by the agent destination

and is accepted by the reputation of the destination

agent then there are two cases. In the ﬁrst case, label

reliability of old knowledge is lesser or equal to the ﬁ-

ability of new knowledge, in which case it updates the

knowledge base of the agent (updateKnowledge). In

the second case, the reliability of the new knowledge

is not superior to what the agent already possessed, in

which case the algorithm allows an alternativemecha-

nism to be used for the acceptance of new knowledge

(conﬁrmKnowledge).

This algorithm provides a basic skeleton for the

processes of acceptance, updating or rejection of

knowledge in an agent. The number and type of la-

bels used can be extended (but may involve increased

complexity of the process).

3.4 Knowledge Spreading in the Agents

Network

We have proposed this model looking to optimize the

number of messages sent through the network to com-

municate knowledge and to establish a structure that

allows some agents to perform certain recommen-

dations or make decisions based on knowledge dis-

persed over different parts of the network. So, our

model proposes the organization of a multiagent sys-

tem in at least three layers. This is in the capabili-

ties knowledge management and communication with

agents (the fact that an agent belongs to one layer or

another will depend primarily on the behaviors that it

implements). ﬁgure 1 shows a diagram of the propa-

gation model.

Figure 1: Model Spreading Knowledge.

As shown in ﬁgure 1, the ﬁrst layer, called the in-

put layer comprises all agents that allow for the en-

try of new knowledge in the system introduced by a

user or generate changes in the knowledge base of the

system from information sent by lower layer agents.

The number of agents that may belong to this layer

is not limited, but the simplest is formed by a sin-

gle agent (decisions and recommendationscentralized

generator). The agents of this layer only communi-

cate with the agents of intermediate layers, but allow

direct communicationwith ﬁnal layer agents when the

response speed requirements are high (ocasionally).

The middle layer comprises all agents that have

the ability to generate decisions or recommendations

based on knowledge acquired from different points or

areas of the multi-agent system and coordinate the

agents belonging to lower layers. This layer can be

subdivided into many middle layers as desired, allow-

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

328

ing for scale and creating a hierarchy multiagent sys-

tem for decision or recommendation generation. The

number of agents in this layer is not limited. It is

usually a number greater than one and less than the

number of agents of the bottom layer that controls. In

the communication level, the agents of this layer are

able to coordinate (horizontal comunication) and dis-

seminate knowledge in the layers in both directions

(upper-layer and lower-layer).

The ﬁnal layer comprises all terminal agents that

have the ability to obtain environmental data or act

on it. The number of agents in this layer is not lim-

ited, it is usually a number greater than the nummbers

in the rest of layers. These agents can be organized

into zones or regions of coverage controlled by a set

of agents from the upper layer. The agents of this

layer only communicate with the agents of the inter-

mediate layers, but allow direct communication with

agents from the input layer when the response time

requirements are high (ocasionally).

ﬁgure 2 shows the different possibilities of the

communication ﬂow modeled, according to the

source of knowledge to be communicated. The possi-

bilities are: (A) Communicationof new knowledge by

a human user, or propagation of an action in the sys-

tem from a global knowledge of system. (B) Commu-

nication of local knowledge: In this case, knowledge

spreads from the ﬁnal agents to agents of the middle

layer or input layer. In the latter case, the spread is

usually done through the middle layer, and it can be

performed directly among agents from the ﬁnal layer

and the input layer (dotted lines) if response time re-

quirements are high. (C) Communication of knowl-

edge to the input layer or ﬁnal layer by an agent of

the middle layer, from a partial knowledge of the sys-

tem.

Figure 2: Spreading Knowledge Flows.

4 PROPOSED ARCHITECTURE

4.1 System Overview

Based on the described model, we propose a general

system architecture composed of four types of agents:

Teleoperator Agent (TA), Coordinator Agent (CA),

Operator Agent (OA) and Device Agent (DA). This

architecture is the basis of the CARISMA project.

The number of coordinator, operator and device

agents is free (speciﬁed in the conﬁguration of the

system), while there is only one teleoperator agent.

Communication among agents is restricted: TA can

communicate with any agent of the system (and vice

versa), while CAs and OAs will have speciﬁc infor-

mation about what other agents they can communi-

cate with. The DAs may only communicate with the

OA to which they are assigned. This conﬁguration al-

lows us to deﬁne ﬂexible areas, by supporting differ-

ent communication channels among agents living in

the system, which can lead to the possibility of over-

lapping in these areas.

An example of the network topology including the

three zones is shown in ﬁgure 3.

Figure 3: General architecture for CARISMA.

The network can expand or shrink according to the

number of the solar power plants and the complex-

ity of the control system. The architecture of indi-

vidual agents is based on the paradigm Belief-Desire-

Intention (Huiliang and Ying, 2005). The Teleopera-

tor Agent is the entry point into the system, providing

a user interface that allows conﬁguration, deployment

and knowledge input to the platform. The Coordi-

nator Agents goals are to coordinate global solutions

for a state of failure or alarm detected from multiple

points in different areas. The Operator Agents are re-

sponsible for controlling the various DAs, and if they

MODEL OF KNOWLEDGE SPREADING FOR MULTI-AGENT SYSTEMS

329

detect failures or local alarms in accordance with in-

formation received from the DAs, they communicate

them to the rest of the system. The Dispositive Agents

are hybrid agents (Cognitive and Reactive), that have

the ability to alert the OA or act directly in case of

changes in the state of a device. Each DA will have a

concrete implementation intended to obtain data from

a particular sensor or perform actions on a given actu-

ator. An example of communication among OAs and

DAs is shown in ﬁgure 4.

Figure 4: Communication architecture for OAs and DAs.

In terms of hardware, there are no restrictions on

the number or type agents that can reside in a device

or style of devices that compose the system, but these

have to be capable of computation. Generally, the

agent node device consists of an embedded system

with support for various transmission technologies

(RF, Ethernet, Bluetooth, ...). One agent node is at-

tached various sensors and actuator devices. The sen-

sors include thermal sensors, humidity sensors, CO2

sensors, sensors for signal from solar plant appliances

and various intelligent meters such as solar irradiation

and video control (Sivianes et al., 2008). The attached

actuators include various valves, motors, and switch-

ers for heating system, ventilation, humidity control,

screen control, etc.

4.2 Application of Model

When applying the model proposed in section 3 to the

architecture discussed in section 4, an automatic con-

trol and decision support system which is very sim-

ple, reliable and scalable can be implemented. For

instance, ﬁgure 5 shows an example of spreading

knowledgein CARISMA, for the B case seen in ﬁgure

2 (generation of knowledge in the ﬁnal layer). OA1,

OA2, OA3 and DA are in the ﬁnal layer of the sys-

tem. CA represents the middle layer and TA of the

input layer.

OA3 does not create any communication, and it is

represented in the ﬁgure as an example of agents be-

Figure 5: Example of spreading knowledge in CARISMA.

longing to the same area but through which no knowl-

edge ﬂows.

The ﬂow of information starts in the DA1 agent,

which communicates the new information acquired

by a sensor to OA1. This fact produces rule activa-

tion, which implies the activation of an alarm to be

communicated to the OA2 in the area. If OA2 has the

knowledge to respond to such an alarm, it then ad-

vises the OA1, otherwise, this alarm must be commu-

nicated to the middle layer, represented by the coor-

dinator agent in its area. Then, if the CA can respond

to the alarm generated in the system, it will commu-

nicate directly to OA1 (in case of action) or to TA (in

case of recomendation). Otherwise it will report the

alarm to the next upper layer which, in this example,

is the input layer, represented by a teleoperator agent.

Finally, if the TA has sufﬁcient knowledge, it can re-

spond directly to OA1 (if imminent action is needed)

or propagate the new information through the system,

by the reverse route. Otherwise, the TA requests in-

formation from the user to respond to the alarm.

Note that information travels in phases through-

out the system and if the necessary knowledge is not

available, then a human is requested. This allows

for the optimization of the information or knowledge

exchanged in the system, and easily increases the

knowledge base of system, with scalability.

5 CONCLUSIONS

Traditionally, control systems have had difﬁculties

as far as their design because they must meet high

requirements. Using our model in the design of a

multiagent-based control system, we can obtain many

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

330

beneﬁts, such as simplicity. Based on the organization

of agents in layers exposed in our knowledge spread-

ing model, we can easily program control systems.

Other problems in control systems are scalability

and ﬂexibility: our model does not limit the number

and types of agents and inclusion / exclusion of layers

is possible. It also allows for dynamic conﬁgurabil-

ity, so we can dynamically change parts of the system

since agents can move from one layer to another, or

new types of agents can be added. It is even possi-

ble to dinamically add new layers with certain restric-

tions, only by changing the behavior of agents. With

regard to the autonomy and intelligence, the agents of

the ﬁnal layer can perform the input and output ac-

tions without the intervention of the central control-

ling computer.

Finally, our model also introduces an optimization

of communication of knowledge among agents in the

system. The layering allows for the design of a hier-

archical system, which leads to a minimization of the

exchanged messages among agents. Additionally, in

the process of acquiring knowledge by an agent, the

model introduced an algorithm that permits the con-

trol of information ﬂows in the system, does so by

using the concepts of reliability and reputation. This

algorithm adds a mechanism to the system that allows

agents to provide solutions or recommendations in a

transparent and intelligent mode.

ACKNOWLEDGEMENTS

The work described in this paper has been funded

by the Consejer´ıa de Innovaci´on, Ciencia y Empre-

sas (Junta de Andaluc´ıa) with reference number P08-

TIC-03862 (CARISMA Project).

REFERENCES

Alessio Lomuscio, M. R. (2000). Intelligent agents vi.

agent theories architectures, and languages. In A Spec-

trum of Modes of Knowledge Sharing between Agents.

Springer, Berlin.

Colomb, R. (2007). Ontology and the Semantic Web. IOS

Press, Nieuwe Hemweg 6B, 1013 BG Amsterdam,

The Netherlands, 1nd edition.

Dongliang, L., Kanyu, Z., and Xiaojing, L. (2008). Eca

rule-based io agent framework for greenhouse control

system. In ISCID ’08. International Symposium on

Computational Intelligence and Design. IEEE Con-

ference Proceeding.

Huget, M.-P. (2003). Communication in Multiagent Sys-

tems: Agent Communication Languages and Conver-

sation Policies. Springer, New York, 1nd edition.

Huiliang, Z. and Ying, H. S. (2005). A parallel bdi agent

architecture. In IEEE/WIC/ACM International Con-

ference on Intelligent Agent Technology. IEEE Con-

ference Proceeding.

Li, H., Karray, F., Basir, O., and Song, I. (2008). A frame-

work for coordinated control of multiagent systems

and its applications. In IEEE Transactions on Systems,

Man and Cybernetics, Part A: Systems and Humans.

IEEE Conference Proceeding.

Mara Adela Grando, C. D. W. (2006). Cooperative infor-

mation agents x. In Specifying Protocols for Knowl-

edge Transfer and Action Restriction in Multiagent

Systems. Springer / Heidelberg, Berlin.

Milner, R. (1994). Communication and Concurrency. Pren-

tice Hall International (UK) Ltd., New York, 2nd edi-

tion.

Sebastian Beer, M. T. (2009). Information technologies

in environmental engineering. In MACE Multiagent

Control for Energy Infrastructures. Springer, Berlin.

Sebestyenova, J. (2005). Decision support system for mod-

elling of systems and control systems design. In In-

ternational Conference on Computational Intelligence

for Modelling, Control and Automation and Interna-

tional Conference on Intelligent Agents, Web Tech-

nologies and Internet Commerce,. IEEE Conference

Proceeding.

Sivianes, F., Romero, M., Hernandez, M., Carrasco, A.,

and Escudero, J. (2008). Automatic surveillance

in power system telecontrol applying embedded and

multi-agent system technologies. In ISIE 2008. IEEE

International Symposium on Industrial Electronics.

IEEE Conference Proceeding.

Walton, C. D. (2003). Multi-agent dialogue protocols. In

Proceedings of the Eighth International Symposium

on Artiﬁcial Intelligence and Mathematics, Fort Laud-

erdale.

Wang Junpu, Chen Hao, X. Y. L. S. (2000). An archi-

tecture of agent-based intelligent control systems. In

Proceedings of the 3rd World Congress on Intelligent

Control and Automation. IEEE Conference Proceed-

ing.

Yang, J., Yang, L., Zhao, T., and Jia, Z. (2009). Automatic

control system of water conservancy project model

based on multi agent. In WKDD 2009. Second Inter-

national Workshop on Knowledge Discovery and Data

Mining. IEEE Conference Proceeding.

Yanping Du, J. Y. (2005). Intelligent agent-based expert

system architecture for generating work plan in mar-

shalling station. In IEEE International Conference on

Granular Computing. IEEE Conference Proceeding.

MODEL OF KNOWLEDGE SPREADING FOR MULTI-AGENT SYSTEMS

331