LOOKING FOR MASCONTROL: A MULTIAGENT SYSTEM

FOR IDENTIFICATION AND CONTROL

E. J. González, Alberto Hamilton, L. Moreno, R. L. Marichal, J.A. Méndez, Vanessa Muñoz

Grupo de Computadoras y Control, Dept. Física Fund. y Exp., Elect. y Sistemas,

Av. Astrofísico Fco. Sánchez S/N, CP. 38207, La Laguna, Spain

Keywords: Identification, control, MAS, ontologies.

Abstract: In this paper, MASCONTROL, a multiagent system (MAS) for system identification and process control, is

presented. This MAS implements a self-tuning regulator (STR) scheme. In this work, an Ontology Agent

(OA) is included, using DAML+OIL as ontology language. From their experience, the authors consider this

architecture highly useful for identification and control processes.

1 INTRODUCTION

Multiagent Systems (MAS) (Faratin et al, 2000)

have been shown to be an effective tool in a wide

range of applications. Nevertheless, researchers

involved in process automation have not taken

significantly advantage of this tool (Cockburn et al,

1996, Velasco et al, 1996, Seilonen et al, 2002).

Some reasons can be found in (Seilonen et al, 2002):

automation applications need real-time requirements

that are out of currently agent technology reach,

difficult agent treatment about interrelations and

lack of parallelism.

The aim of this paper consists of presenting the

development of a multiagent system oriented to

process automation. The development of this

standarized and complex multiagent system, called

MASCONTROL is detailed in Sections 2-6: a

preliminary introduction, a justification of the

selected tools, description of the designed

framework and the designed ontology (this

multiagent system includes an ontology agent for

communication with external agents) and results

obtained in the control process of a real plant.

Finally, conclusions are reported in Section 7.

2 INTRODUCTION TO

MASCONTROL

FIPA (Foundation for Intelligent Physical Agents)

specifications have become a stronger standard in

MAS development and they involve not only agent

language specifications but also agent management,

conversations, etc. The FIPA agent management

reference model provides the normative framework

within which FIPA agents exist and operate. The

Directory Facilitator (DF) provides yellow pages

services to agents that query it to find out services

offered by other agents. The Agent Management

System (AMS) offers white pages services and

maintains a directory, which contains transport

addresses for agents registered in the Agent Platform

(AP). The Message Transport Service (MTS) is the

default communication among agents on different

APs.

In this paper, the paper have applied the power of

the MAS to control disciplines. In particular, the

involved agents are supposed to implement the

different roles related to a self-tuning regulator, STR

(Aström and Wittenmark, 1995, Söderström and

Stoica, 1989, Warwick, 1989). The authors consider

this problem as a very interesting one in the

application of MAS. The first reason is that a STR

presents conceptually different modules, such as

sampling, identification or control. Other reason

consists of the parallelism in the system, optimizing

the calculation capacity this way. Both reasons make

MAS appropriate.

3 SELECTION OF THE TOOLS

This section justifies briefly the selection of the used

tools for the MAS development. It is pretended to be

119

J. González E., Hamilton A., Moreno L., L. Marichal R., A. Méndez J. and Muñoz V. (2005).

LOOKING FOR MASCONTROL: A MULTIAGENT SYSTEM FOR IDENTIFICATION AND CONTROL.

In Proceedings of the Second International Conference on Informatics in Control, Automation and Robotics, pages 119-124

DOI: 10.5220/0001173401190124

 SciTePress

an elementary survey guide for those researchers

that would like to use a MAS in their projects.

3.1 Multiagent framework

Once FIPA standard has been chosen for the

development of the system, the following step

consists of the selection of the multiagent

framework. This fact involves being not worried

about technical aspects such as a detailed

communication implementation. With regard to this,

there are many tools for the design and

implementation of MAS.

In this sense, the authors have chosen FIPA-OS.

This is an Open Source agent framework from

research at Nortel Network’s Harlow Laboratories

that implements the FIPA specifications about agent

interoperability. There are other Java-based

implementations of FIPA specifications such as

JADE and ZEUS. However, FIPA-OS has been

chosen in this work due to

its well-placed use of Java interfaces to separate

agent subsystems, translation of incoming

messages and system occurrences into events for

internal processing, an isolated scheduling policy

for task execution, use of a conversation object

to enforce protocols and hold messages, and a

task generation tool for constructing tasks from

protocol definitions (Fonseca, 2002)

3.2 Ontology Language

There are few situations where a standalone agent is

able to manage an open and dynamic environment:

agents are designed for communicating each other.

In this sense, having a shared ontology (Falasconi et

al, 1996) makes a multiagent system more powerful,

as agents developed by different designers can

interact in the same environment. In this case, the

problem to solve is too easy (only few terms in the

ontology are necessary for the problem, as can be

seen below). However, it is interesting to consider

its use in more complex situations.

For ontology design, there are many possible

ontology languages. The authors have used

DAML+OIL, a markup language, as ontology

language, because it provides a basic infrastructure

that allows a machine to make some sorts of simple

inferences that human beings do and because this

language has become a standard in ontology

representations. At this moment a new, improved

and more expressive markup language, known as

OWL, is being developed. Nevertheless

DAML+OIL is sufficiently expressive for carrying

out this work.

This selection involves the use of some related

software, in particular an ontology editor (OilEd has

been chosen) and a DAML+OIL parser (Jena).

3.3 Evenet2000

The problem of optimization is much related to

identification and control. In both disciplines, it is

usually necessary to optimize some parameters for a

model (identification) or for a controller. In the same

way, the training of a neural network consists of

finding the best values of the weights of the

network. Because of this similarity of methods, the

authors have considered the possibility of the

application of neural networks training methods to

control problems.

For this MAS, the authors have preferred used

Evenet2000 (Gonzalez et al, 2001 and 2003), a Java-

based neural network toolbox developed at the

University of La Laguna. This Open-source tool

offers some features such as the possibility of

designing an arbitrary neural network to optimize.

This tool develops an approach to derive gradient

algorithms for time-dependent neural networks,

using the Signal Flow Graph theory. Moreover, the

designed structure makes it not to be limited to

algorithms based on the gradient. This tool has been

shown as highly useful in a suite of control

problems.

In MASCONTROL, Evenet2000 is used in two

ways. Firstly, for the optimization of the parameters

of a model. Secondly, in the optimization of the

parameters of a controller such as the value of K

a proportional controller. The different agents import

Evenet2000 packages for its use (Gonzalez et al,

2004).

4 MASCONTROL AGENT

FRAMEWORK

This framework is composed of 8 different types of

agents. Each agent assumes one of the roles that

appear in a STR scheme.

ReaderCommandAgent (RCA) This agent samples

the output of the system. Another role assumed by

this agent consists of calculating and sending the

command to the system. Finally, it stores a vector

containing input-output data. This vector makes

ICINCO 2005 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION

120

possible the system identification and determines if

the input is rich enough for that identification.

IdentificationAgent (IA) Several identification

agents can appear in the system. Each IA tries to

identify the system from the input-output vector. For

this purpose, it uses Evenet2000 modules. A system

user or an IdentificationLoaderAgent (from a record

of previous trainings) can select a training method

for each IA. The system model is declared in the

XML profile and loaded as an Evenet2000 neural

network file. In other words, for an IA, the problem

is equivalent to the one of weight optimization in a

neural network whose training pattern is defined by

the input-output vector. The authors have included a

one-step ahead predictor in the identification

process.

LinearIdentificationAgent (LIA) Similar to the IA,

but this agent assumes a model that allows a linear

regression, for example an ARX model. In this

sense, the model is defined by the orders of the

numerator and the denominator of the transfer

function. Instead of Evenet2000 modules, an object

that implements an identification through the

Forgetting Factor technique is used (Aström and

Wittenmark, 1995, Söderström and Stoica, 1989,

Warwick, 1989).

CentralIdentificationAgent (CIA) This agent

manages the Ias, linear or not. Initially it asks the DF

for the agents that offer the service of identification

in the MAS. Every T seconds, the CIA asks the IAs

for the results of the current optimizations (error,

parameter set and model file), selecting that

optimization offering the best results. Then it

informs the rest of the IAs with the same model the

set of parameters that minimize the criterion

function. This way, the IAs take this set as an initial

training one in new optimization processes. With the

aim of providing the MAS with some intelligence,

the CIA counts how many times each training

method obtains the best results optimizing the

criterion function. This information is stored in a file

(written in DAML+OIL), that could be used in the

initiatization other IAs.

OptimizerAgent (OpA) This agent optimizes the

controller parameters. For this purpose, it takes the

set calculated by the identification agents and

includes it as constants in a new system. This system

is treated as a neural network whose parameters are

the controller ones. In a general way, patterns are

chosen as pairs reference input, reference input in a

serie of different reference inputs. With this

implementation, high raising time and valued peaks

are penalized. The model of the system can be easily

changed due to Evenet2000 modularity.

CentralControlAgent (CCA) This agent is similar

to CIA as it plays a manager role in the system, but

in this case it is related to the optimization of model

parameters. Each T seconds, the CCA asks the CIA

for the details (model with the best results,

parameter set) of the identification. After analyzing

these data, the CCA asks the OpAs for the

parameters that minimize the criterion function and

for the value of this minimization. This agent stores

the results for subsequent sending to the RCA.

Finally, the CCA orders the OpAs to stop the current

optimizations and to start a new one from the

calculated optimal parameter set. As the CIA, the

CCA stores the number of times that a given training

method has offered the best results for the analyzed

control process.

InputOutputAnalyzerAgent (IOAA) This optional

agent analyzes process input and output data

(calculated by the RCA). This analysis is made in

two ways. First, it tests, in an intuitive way, if the

system input is rich enough. For this purpose, this

agent calculates the maximum and minimum input

value in the last N periods, and it tests if the

subtraction of these values is less than a defined low

enough threshold. If it is the case, it is supposed that

the applied input is not rich enough and the IOAA

suggests the RCA that the reference input should be

changed. This change is supposed to improve the

identification process. In a similar way, output data

are analyzed too. In this case, the IOAA could

suggest a reference input change or a study about the

type of the system. The option of changing the

reference input can be inhibited on-line through a

user interface.

Ontology Agent (OA) This is one of the key

considerations in this phase of the work. This way, it

differs from other MAS-based control systems. The

definition of an external ontology provides

numerous advantages: it allows consultation with

respect to concepts, the updating and use of

ontologies and the elimination of the need of

programming the entire ontology in every agent,

hence reducing required resources. Currently, in

MASCONTROL, this agent only takes part in the

study of the type of the system.

5 MASCONTROL ONTOLOGY

This section will describe the main features defined

in the implemented ontology. As stated above, most

of the concepts have not been used in

MASCONTROL. However, they are defined

looking for an open system where a new agent,

LOOKING FOR MASCONTROL: A MULTIAGENT SYSTEM FOR IDENTIFICATION AND CONTROL

121

implemented by other researchers, could interact

with MASCONTROL agents. Defined classes are

mainly related to control concepts, such as System,

Input, Output, ReferenceInput, Error…

A second group of classes are referred to the

implemented optimization algorithms in

Evenet2000: OptimizationMethod, TrainingMethod.

As an example, the following DAML+OIL code

defines NonLinearSystem as a subclass of System.

<daml:Class rdf:about="#NonLinearSystem">

<rdfs:subClassOf>

<daml:Class rdf:about="#System"/>

</rdfs:subClassOf>

</daml:Class>

With respect to the defined properties, these are

mainly referred to the values of defined classes:

valueOfPole, valueOfZero. On the other hand, other

properties are related to control concepts such as

order and type of a system. These definitions allow

the system to make some interesting inferences from

some axioms defined in the ontology. For example,

if a given system does not reach a desired reference

input with a proportional control action, the MAS

could deduce that the type of the system is 0. A

suitable control action for a system like that is a

proportional-integral (PI) one. Therefore, an agent

can deduce that a PI control action is a suitable

control action for a system that does not reach a

desired reference input with a proportional control

action and when this axiom does not appear

explicitly in the ontology.

6 RESULTS

With the proposed agent architecture, several

experiments related to the control of a plant have

been carried out.

6.1 Description of the plant

The plant consists of an educational module for the

control of processes. This module, shown in the

Figure 1, is from the PROCON (Process Control

Trainer) series of the Feedback Company. for the

level and flow control.

The plant consists of a closed-circuit of water that

allows the study of the principles of control

considering the water level and flow as variables to

control. The plant includes a dual tank in the top

where water level is measured. This tank is

connected to a store-tank through five manual and

three solenoid valves. Water is pumped from the

store-tank to the dual tank through a flowmeter and a

servovalve. The plant is controlled through a

standard 4-20 mA action, although for

implementation purposes the reference values (and

the outputs) are set/measured in volts. This

limitation makes more difficult the optimization of

the controller parameters, introducing a non-

derivability in the Criterion Function. This fact is

perfectly managed by the modules designed in

Evenet2000 although it can be defined a Criterion

Function that would punish those values out of the

allowed ones.

Figure 1: Plant controlled by MASCONTROL

6.2 Experiments

The experiments were carried out using three

different control actions: proportional, proportional-

integral and pole placement. The identification

process has been carried out through several IAs

acting simultaneously in the multiagent system.

These agents have used several models for the plant,

both linear and not linear. Moreover, the

identification and the optimization of the parameters

of the controller have been carried out using

simultaneously different optimization methods. With

an initial value k

=1, a learning phase is carried. In

this phase reference input value is continually

modified, looking for a better identification. When it

is considered that identification is good enough

(after a long time, when parameters values are

stabilized or manually through a user-interface),

reference input is set to the desired value.

ICINCO 2005 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION

122

In the following sub-subsections the authors will

analyze the results related to the PI and pole

placement control actions.

6.2.1 PI Control Action

Figure 2 shows the optimization carried out by the

OpAs. with a variable reference (3, 1, 2, 0.8 and 1.5

V.). In this sense, with this type of sequence

(increasing/decreasing), the model behaviour is

supposed to be more independent from a given

reference input.

In this case, it can be seen that the system output

reaches the reference input in each section, without

overshoot. This optimization is reflected in the

system output: it reaches the desired reference inputs

(2.5, 3.7, 4 and 3 V) and with a low overshoot. The

two mentioned phases (training and desired

reference input) can be distinguished (see Figure 3).

Moreover, the measured constant time (150 seconds

aprox.) is of the same order as the measured one.

This fact is considered to be a validation of the

MAS.

Figure 2: Optimization result obtained by the OpAs

Figure 3: Results obtained using PI action control

The following experiment illustrates the use of an

OA, and ontologies in general, in a control system.

Figure 3 shows the output evolution initially

controlled by a proportional control action. When

the IOAA realizes that the system output is

stabilized and that it has not reached the reference

input, the RCA asks the OA for the suitable control

action when a system does not reach a reference

input using a proportional action. At this point, the

OA looks for this fact in the system ontology and

informs the RCA that the answer is a proportional-

integral control action. Then, MASCONTROL

makes the necessary modifications (control action

and model). As it was expected, the process output

reaches the desired reference input after these

modifications. This fact is only an example of the

power that ontologies are able to supply to a process

control system.

Figure 4: Ontology Effect: P to PI control action

6.2.2 Pole replacement control action

The following work focuses on studying that

goodness for a totally different control action, in

particular, pole replacement. This control action

depends critically on system identification.

Figure 5: Pole replacement control action

Due to MASCONTROL modularity, some minor

modifications are necessary. Firstly, the RCA should

implement this new control action. Secondly,

controller parameter set optimization has no sense

for this control action. Therefore, the CCA is only

limited to propagate CIA results. Figure 5 shows the

obtained results. In this case, as it can be seen, the

system identification (with proportional-integral

control action) phase has been longer, looking for a

better identification.

LOOKING FOR MASCONTROL: A MULTIAGENT SYSTEM FOR IDENTIFICATION AND CONTROL

123

7 CONCLUSIONS

The development of a MAS architecture for process

control is presented in this paper. In particular this

MAS, called MASCONTROL implements a self-

tuning regulator (STR) scheme. It is another

alternative to solve problems of process control.

This MAS has adopted FIPA specifications because

they have become a stronger standard in MAS

development and it involves not only agent language

specifications but also agent management and

conversations.

In this work, an Ontology Agent (OA) is included.

An OA is an agent that provides access to one or

more ontology servers and which provide ontology

services to an agent community. For ontology

design, DAML+OIL is a sufficiently expressive

language for carrying out this work. Apart from the

standard tools, the authors have included

Evenet2000. This toolkit has been shown to be

highly useful in control problems.

The authors strongly recommend the use of MAS

and ontologies in order to develop applications such

as those analyzed in this paper. The results that are

reported in this paper indicete good future

possibilities. They expect to have shown that its use

offers more than they cost.

REFERENCES

Aström K.J. and Wittenmark B., 1995, Adaptive Control -

Second Edition. Addison-Wesley (ISBN 0-201-55866-

1).

Cockburn D. and Jennings. N.R., 1996, ARCHON: A

Distributed Artificial Intelligence System for Industrial

Applications. In Foundations of Distributed Artificial

Intelligence, G. M. P. O'Hare and N. R. Jennings,

editors.

Falasconi S., Lanzola G., Stefanelli M., 1996, Using

ontologies in Multi-Agent Systems. Tenth Knowledge

Acquisition for Knowledge-Based Systems Workshop,

1996

Faratin P.; Jennings N.R.; Buckle P. and Sierra C., 2000.

Automated Negotiation for Provisioning Virtual

Private Networks using FIPA-Compliant Agents. Proc.

5th Int. Conf. on the Pratcical Application of

Intelligent Agents and Multi- Agent Systems (PAAM-

2000), Manchester, UK, 185-202

Fonseca S.P., Griss M.L., Letsinger R. Agent Behavior

Architectures. A MAS Framework Comparison.

Proceedings of the first international joint conference

on Autonomous agents and multiagent systems,

Bologna, Italy, 2002.

González E.J. , Hamilton A., Moreno L., Marichal R.L.,

Muñoz V., 2004, MASCONTROL: A MAS for System

Identification and Process Control, 8th Pacific Rim

International Conference on Artificial Intelligence,

Auckland, New Zealand, August 9-13.

Gonzalez E.J., Hamilton A., Moreno L., Aguilar R.M.,

Marichal R.L., 2003, Neural networks teaching using

Evenet-2000. Computer Applications in Engineering

Education. Volume 11, Issue 1, 1-5.

Gonzalez E.J., Hamilton A., Moreno L., Sigut J., Marichal

R.L., 2001, Evenet-2000: Designing and Training

Arbitrary Neural Networks in Java. Bio-Inspired

Applications of Conectionism. Springer-Verlag (2085)

104-110. 2001.

González E.J., Marichal R.L., Hamilton A.F., Moreno L.,

2003, Evenet2000 as Software Application for System

Identification and Control. The 11th Mediterranean

Conference on Control and Automation. Rhodes.

Greece.

Seilonen I., Appelqvist P., Vainio M., Halme A.and

Koskinen K., 2002, A Concept of an Agent-Augmented

Process Automation System. 15th IFAC World

Congress on Automatic Control, Barcelona, Spain.

Söderström T., Stoica P., 1989, System Identification.

Prentice Hall International Series in Systems and

Control Engineering.

Velasco J., González J.C., Magdalena L. and Iglesias C.,

1996, Multiagent-based Control Systems: a Hybrid

Approach to Distributed Process. Control Engineering

Practice. Vol. 4 839-846.

Warwick K., 1989, Control Systems: An Introduction.

Prentice Hall International Series in Systems and

Control Engineering.

ICINCO 2005 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION

124