Self-organizing Agents for an Adaptive Control of Heat Engines

emy Boes, Fr

eric Migeon and Franc¸ois Gatto

Institut de Recherche en Informatique de Toulouse, Universit

e Paul Sabatier, Toulouse, France

Keywords:

Multi-Agent Systems, Self-organization, Intelligent Control Systems.

Abstract:

Controlling heat engines imposes to deal with high dynamics, non-linearity and multiple interdependencies.

A way handle these difﬁculties is enable the controller to learn how the engine behaves, hence avoiding the

costly use of an explicit model of the process. Adaptive Multi-Agent Systems (AMAS) are able to learn and

to adapt themselves to their environment thanks to the cooperative self-organization of their agents. A change

in the organization of the agents results in a change of the emergent function. Thus we assume that AMAS

are a good alternative for complex systems control, reuniting learning, adaptivity, robustness and genericity.

In this paper, we present an AMAS for the control of heat engines and show several results.

1 INTRODUCTION

Over the years, the complexity of car engines has been

growing up, with new sophistications like Exhaust

Gas Recirculation or particle ﬁlters. Moreover our ex-

pectations about their efﬁciency are becoming more

important due to the strengthening of environmen-

tal norms, while the need for economic competitivity

leads to a need for easily instantiable controllers. The

emergence of software and electronic devices helps

to tackle these challenges but the development of En-

gine Control Unit (ECU) remains a difﬁcult and time

consuming task.

The goal of this paper is to introduce a non-

classical approach of the control of car engines, based

on Adaptive Multi-Agent Systems (AMAS), that aims

to provide a quickly instantiable controller, avoiding

a heavy parametrization work. Based on learning and

adaptiveness skills of AMASs, our approach consid-

ers the engine and its ECU as a black box. The con-

troller modiﬁes ECU parameters in order to obtain the

desired values on the engine outputs. This work is

part of a national project named Orianne.

Our motivation for working on a non-classical ap-

proach comes from the difﬁculties to build and ﬁt an

efﬁcient model of the engine we intend to control.

Humans fulﬁll this control task in their everyday life.

A child learning to drive a bike is a good example.

He has the ability to move his legs in order to walk

and run. He does not have any knowledge about bike

gears or inertia theory. While he is experiencing his

ﬁrst ride, his most used skill is the ability to adapt and

to coordinate his movements in a new situation. This

adaptation requires learning, which is certainly not an

instantaneous process. Following a practice period,

the child eventually ends up acquiring the ability to

control the bike but still did not get in touch with any

theoretical concept and did not build any analytical

model of the bike inside his head. Furthermore, he

never stops learning: each new ride will improve his

mastering of the bike, and he will be able to adapt to a

new bike. This analogy is meant to explain the point

of view we adopted: the use of an explicit analytical

model is not necessary to the control of an engine.

Section 2 introduces usual control methods. We

follow with an introduction to self-organization and

to the Adaptive Multi-Agent Systems (AMAS) the-

ory. Section 3 describes the design of a Multi-Agent

System before going deeper in the agents behavior in

section 4. Then section 5 shows some results, before

section 6 concludes with our perspectives.

2 RELATED WORKS

In this section the main approaches of control are

presented before self-organization and the Adaptive

Multi-Agent Systems theory are brieﬂy introduced.

2.1 Complex Systems Control

Controlling systems is a generic problem that can be

expressed as ﬁnding which modiﬁcations are needed

to be applied on the inputs in order to obtain the de-

243

Boes J., Migeon F. and Gatto F..

Self-organizing Agents for an Adaptive Control of Heat Engines.

DOI: 10.5220/0004483302430250

In Proceedings of the 10th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2013), pages 243-250

ISBN: 978-989-8565-70-9

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

sired effects on the ouputs. The main approaches are

presented in the next paragraphs.

2.1.1 PID

Proportional-Integral-Derivative (PID) controllers are

well-known and widely used. They base the control

on the calculation and the combination of three terms,

corresponding to the error (P), its duration (I) and its

variation rate (D) (Astrom and Hagglund, 1995). The

instantiation of a PID controller to a particular system

is a heavy procedure despite adjustement methods

like (Tyreus and Luyben, 1992). Besides, PID con-

trollers are not always efﬁcient with non-linear sys-

tems, they need to be complemented with other mech-

anisms like fuzzy logic (Visioli, 2001). Also, a PID

controller can only handle one input with one retroac-

tion signal, which is a severe drawback for complex

systems control.

2.1.2 Adaptive Control

The dynamic nature of complex systems imposes to

adapt the controller as the system evolves. Model-

based approaches like Model Identiﬁcation Adaptive

Control (MIAC) (Soderstrom and Stoica, 1988) or

Model Reference Adaptive Control (MRAC) (Kreis-

selmeier and Anderson, 1986) are regrouped under

the name of Model Predictive Control (MPC) (Niko-

laou, 2001). A model able to forecast the behavior of

the process is used in order to help the controller ﬁnd-

ing the optimal control scheme. These approaches

successfully handle several inputs of dynamic non-

linear systems. They are used in particular in the do-

main of car engines (Dabo et al., 2008) but are limited

by the construction of the model on one hand, and

its parametrization on the other hand, which can take

several years.

2.1.3 Intelligent Control

Intelligent control can be seen as a part of adaptive

control. It regroups control approaches that use Ar-

tiﬁcial Intelligence methods to enhance the controller

adaptiveness. Among these methods we can ﬁnd neu-

ral networks (Hagan et al., 2002), fuzzy logic (Lee,

1990), expert systems (Stengel, 1991) and bayesian

controllers (Colosimo and Del Castillo, 2007). Neu-

ral networks provide great genericity but they require

an ofﬂine learning step with a considerable amount

of data. Even if this enable an efﬁcient control, they

remain sensible to perturbations and cannot adjust

themselves to the evolution of the system.

Expert systems are composed of a knowledge

base and an inference engine. They lack of generic-

ity and are hardly adaptable to highly dynamic non-

linear systems. That is why they are often com-

bined with fuzzy logic. Bayesian controllers are

based on bayesian probabilities, the most common are

the Kaman ﬁlter based controllers (Lee and Ricker,

1993). Yet, the bayesian approach does not deal with

the problem of control itself but with a sub-problem

which is the observation of the process. These meth-

ods can also be easily combined one with another, or

with more classical control approaches.

Finally agents-based techniques have been applied

in various ﬁelds, for example in the case of voltage

control (Wang, 2001) or manufacturing control (Clair

et al., 2008).

2.2 Self-organizing Software

The next paragraphs introduce the notion of self-

organization along with the Adaptive Multi-Agent

Systems theory.

2.2.1 Self-organization

The notion of self-organization refers to the fact that

the organization of a system results from internal

mechanisms and local interactions without any ex-

plicit external control (Camazine et al., 2003). Var-

ious artiﬁcial systems show this property, such as

swarms (Montresor et al., 2003) or some multi-

agent systems (O’Hare and Jennings, 1996). Self-

organization brings robustness and scalability, and

allows the engineering of underspeciﬁed software

(Di Marzo Serugendo et al., 2011).

2.2.2 Adaptive Multi-Agent Systems

The Adaptive Multi-Agent Systems (AMAS) the-

ory is a basis for the design of multi-agent systems

where cooperation is the engine for self-organization

(Georg

e et al., 2011). As cooperative entities, AMAS

agents try to reach their own goals as well as they

try to help other agents to achieve their goals. More-

over, an agent will modify its behavior if it thinks that

its actions are useless or detrimental to its environ-

ment. Such situations are called Non-Cooperative Sit-

uations (NCS). Some behavioral rules are speciﬁc to

NCSs, they allow agents to avoid or to solve these sit-

uations. By solving NCSs in regard to their own local

goals, cooperative agents collectively ﬁnd a solution

to the global problem. Therefore one can consider the

global behavior of an AMAS as emergent.

AMAS principles have been applied by (Videau

et al., 2011) for the control of bioprocesses, avoid-

ing the use of a speciﬁc model and dealing with the

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

244

Figure 1: Criticality Functions.

evolution of the process over time, since an AMAS is

perpetually adjusting itself to its environment.

The next two sections describes the agents of our

controller, how they perform control and how they

self-organize.

3 CONTROLLER OVERVIEW

In this section we present the required abilities of a

complex system controller, and what are the agents of

our controller that enable them.

3.1 Basic Behavior

The next paragraphs describe how our multi-agent

system, called ESCHER (Emergent Self-adaptive

Controller for Heat Engine Regulation), works when

it is already adapted to the engine. The mechanisms

that lead to this adaptation will be explained further.

3.1.1 Observing the Process

If we intend to control an engine, it is obvious that we

need to be able to observe it. A speciﬁc agent type is

in charge of perception, called Variable Agent. Each

input and each output of the controlled black box has

a corresponding Variable Agent. These agents per-

ceive their value from the black box and send it to

other agents that need this information. Also, Vari-

able Agents can embed noise reduction algorithms if

this problem is not handled by a third party system.

3.1.2 Representing Objectives and Constraints

The controller needs to know what is the desired state

of the engine. This state is represented by a set of

Constraint Agents and possibly by additional Variable

Agents.

There are three types of Constraint Agents:

Threshold, Setpoint and Optimization. A Threshold

Constraint Agent expresses the will to keep a vari-

able either below or above a threshold speciﬁed by

a Variable Agent. This Variable Agent can either

correspond to a process variable or to a user-deﬁned

value. In a similar way, a Setpoint Constraint Agent

Figure 2: ESCHER and its Environment.

expresses the will to set a process variable to a partic-

ular value speciﬁed by a Variable Agent. Finally, an

Optimization Constraint Agent represents the will to

minimize or maximize a process variable.

Using the functions shown in ﬁgure 1, each Con-

straint Agent computes a critical level that varies from

0 (the agent is satisﬁed) to 100 (the agent is far from

satisﬁed). The critical level depends on the value of

the variable on which the constraint is applied and

can be parametrized by a second Variable (in the case

of a threshold or a setpoint constraint). The critical-

ity function of an Optimization Constraint Agent is

asymptotic to zero, since we always try to minimize or

maximize the value. For instance, if the user desires

to set the rotational speed to 5000rmp, the Variable

Agent RotSpeedCommand (corresponding to “user-

deﬁned setpoint on rotational speed”) will be set to

5000 and a setpoint Constraint Agent will be created.

Its critical level will be zero if the value of the Vari-

able Agent RotSpeed equals the value of RotSpeed-

Command, and will increase with the difference be-

tween RotSpeed and RotSpeedCommand.

It is clear that decreasing the critical levels means

solving the constraints, and the only way to do so is

to perform the adequate actions on the engine inputs.

Finding these actions implies to be able to analyze the

current sate of the environment. The engine and the

user desires are considered as the environment of the

multi-agent system.

3.1.3 Analyzing the State of the Environment

As shown in Figure 2, ESCHER environment is the

engine as well as the user-deﬁned setpoints, thresh-

olds and optimizations. Thanks to Variable Agents

and Constraints Agents, the ESCHER has a represen-

tation of its environment. Before it can perform con-

trol, it must be able to extract relevant information

from this representation. This is the role of agents

called Context Agents.

A Context Agent memorizes the effects, on ev-

ery critical level, of an action applied to one partic-

ular input of the engine. It also memorizes the state

of the environment when the action was applied. To

represent this state the Context Agent maintains a set

Self-organizingAgentsforanAdaptiveControlofHeatEngines

245

of validity ranges, containing one range per Variable

Agent. A Context Agent also maintains a set of values

that are the observed and memorized effects of the ac-

tion on each of the critical levels. This set is called the

forecasts. In other words, a Context Agent represents

the information that if every Variable Agent value is

inside the Context Agent validity range, and if its ac-

tion is applied, then the effects on every critical level

will be similar to the Context Agent forecasts.

A Context Agent is said valid when the environ-

ment is in a state that matches its validity ranges, i.e.

every perceived variable is inside the corresponding

validity.

When valid, a Context Agent sends a notiﬁcation

with its action and forecasts to the appropriate Con-

troller Agent, which will be explained in the next part.

3.1.4 Selecting the Adequate Action

Each controlled input of the engine is associated with

a Controller Agent. The role of a Controller Agent is

to apply the most adequate action in order to reduce

the critical levels. It will base the choice of the action

(ie increment, decrement or stay put) on the informa-

tion it receives from Context Agents, picking the ac-

tion that will provoke the biggest decrease of the crit-

ical levels. When an action is picked, the Controller

Agent notiﬁes every Context Agents who proposed it.

For instance, the Controller Agent of the mass fuel

injected may receive two notiﬁcations from two Con-

text Agents. Context Agent A says if you apply a

decrease of 0.5 the critical levels will not change ex-

cept for the Constraint Agent Torque Optimization

that will rise of 3.5. Context Agent B says if you apply

an increase of 0.5 the critical levels will not change

except for the Constraint Agent Torque Optimization

that will drop of 1.5. In this case, the Controller Agent

obviously chooses to increase the mass fuel injected.

There are several cases where the Controller

Agent is unable to make a good decision, because

of incomplete or incorrect information from Con-

text Agents. These cases are Non-Cooperative Situa-

tions (NCS). They happen when ESCHER is not well

adapted to the engine yet. When a NCS occurs, the

cooperative behavior of involved agents is triggered

in order to solve it. This will be explained in section

4. For now, we continue with the architecture of the

system.

3.2 Topology

The four ESCHER’s agent types were introduced in

the previous paragraphs. Figure 3 shows the topol-

ogy of their interactions. We can see that, while it is a

MAS itself, ESCHER is also an aggregation of several

MASs, each constituted by a Controller Agent and its

set of Context Agents. The environment of these sub-

MASs is composed of the Variable Agents, the Con-

straint Agents, and the environment of ESCHER. In-

deed, each Controller Agent is not directly in interac-

tion with other Controller Agents. The only link be-

tween them are their environment (the engine they are

controlling, via the Variable and Constraint Agents).

Actions from one Controller Agent will have conse-

quences on the engine that will be reﬂected by Vari-

able and Constraint Agents, and Controller Agents

share the same environment, hence we can expect

them to progressively synchronize their actions. Fur-

thermore, Context Agents never directly interact with

each other (even inside the same set) but only with

their associated Controller Agent.

The question is now how can the system create

and adjust Context Agents in order to deal with an

unknown process? The next section gives an answer.

4 CONTEXT AGENTS LEARNING

AND ORGANIZATION

This section explains how the set of Contexts Agents

is built and how they self-organize to provide an accu-

rate representation of the explored part of the process’

state space.

4.1 Non-cooperative Situations

In this part, we explain how agents solve the Non-

Cooperative Situations they face. Since they provoke

changes in the agents behavior, NCSs are the key of

the adaptiveness of Adaptive Multi-Agent Systems.

Each agent locally solves its own NCSs when they

occur, by changing its own local behavior, and thus

of the global behavior of the system. The NCSs of

ESCHER concern the Controller Agents and the Con-

text Agents, and trigger behaviors that correct Context

Agents’ forecasts and validity ranges.

4.1.1 No Adequate Action in Suggestions

This NCS occurs when the suggestions list of a Con-

troller Agent contains only forecasts of increasing

critical levels. Whatever the Controller Agent may

chose, it believes it will deteriorate the constraints.

There are two cases: either all the possible actions are

already suggested (i.e. increase the value, decreasethe

value and stay put) or some actions are not suggested.

In the ﬁrst case, the only choice left is to accept the

suggestion with the less bad forecasts. In the second

case, a new action is chosen among those which are

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

246

Figure 3: ESCHER Topology.

not suggested, and a new Context is created with this

action.

4.1.2 Empty Suggestions List

This NCS happens when a Controller Agent has to

apply an action, but ﬁnds its suggestion list empty.

It will be unable to ﬁnd an adequate action with cer-

tainty, but it can make some hypothesis to try one. If

the last applied action had reduced the higher critical

level, the same action is reproduced. If not, the op-

posite action is applied (e.g. if the last action was an

inscrease of the input value, the new action will be a

decrease). If no action has ever been applied, the so-

lution is to randomly chose the action. In any case a

new Context Agent is created with the applied action.

After its creation, a Context Agent extends its validity

ranges as long as its action is applied and observe the

variations of the critical levels to set its forecasts.

4.1.3 Wrong Forecast

When a Context Agent is selected and its action ap-

plied, it oberves the variations of the critical levels to

check if they matches its forecasts. If the forecasts

are wrong (i.e. a critical level evolves in the opposite

direction of the forecast), the Context Agent consid-

ers that it should not have been valid when the ac-

tion was suggested. Thus it reduces its validity ranges

but does not modify its forecasts. A Context Agent

dies if one of its ranges is reduced to an amplitude of

zero. If the forecasts are simply inexact (i.e all the

critical levels evolve in the forecasted direction, but

not with the forecasted amplitude), the Context Agent

considers that its validity was relevant. Thus it does

not modify its validity ranges but adjusts its forecasts

to match its observation.

This NCS also impacts the Controller Agent that

had selected the Context Agent with the wrong fore-

casts. When a Controller Agent notices that the

critical levels variations don’t match the forecast, it

stop applying the current action, notiﬁes the Context

Agent that its action is no longer applied, and looks

for a new action in its suggestion list.

4.2 Self-organization

Context Agents are able to evaluate their own behav-

ior and to adjust it (by modifying their forecasts and

their validity ranges) if necessary. Thus we can say

that a Context agent is self-adaptive. Moreover, Con-

text Agents are created at runtime. Each of them fol-

lows simple and local behavioral rules. These rules

lead to the formation of a coherent set of Context

Agents where each agent occupies a portion of the en-

vironment state space for which its forecasts on crit-

ical levels are correct. Thus we can say that the set

is the result of the self-organization of the Context

Agents. It is also intersting to note that all the Con-

text Agents of a given set put together give a good

approximation of the criticality functions.

There is another, more subtle, level of self-

organization. Each Controller Agent makes its own

decisions about the action to apply on its effec-

tor, other Controller Agents are never consulted.

Nonetheless, it appears that they manage to synchro-

Self-organizingAgentsforanAdaptiveControlofHeatEngines

247

nize their actions and efﬁciently reduce the critical

levels, solving the constraints. This is possible be-

cause each process input corresponding to a Con-

troller Agent is also represented as a Variable Agent.

Context Agents suggestions for one Controller Agent

are therefore conditionned by the current state of

other Controller Agents corresponding process input.

As they all try to lessen the critical levels, they even-

tually ﬁnd a synergy and the global behavior of ES-

CHER (the actions on all the controlled inputs) is co-

herent.

5 RESULTS

Generated black-boxes (Boes et al., 2013) were used

for the ﬁrst tests to check ESCHER’s abilities to deal

with multiple inputs and outputs, to ﬁnd compro-

mises between several objectives, and to adapt to un-

expected changes of behavior of the controlled pro-

cess. These tests are not detailed here due to space

limitations.

Several tests have been driven on a 125cc mono-

cylinder fuel engine in various steady operating points

of rotational speed and load. Those presented here

have been driven at 5000rpm and 870mbar of pressure

in the inlet manifold. The goal of ESCHER is here

to maximize the indicated mean effective pressure

(IMEP), an output variable that reﬂects the torque. In

other words, the goal is to minimize the critical level

of an Optimization Constraint Agent associated to the

IMEP. Injected mass fuel (IMF) and ignition advance

are controlled by ESCHER, which has no other pre-

vious information about the engine than the variation

interval of the variables. Each test has been super-

vised and validated by domain experts.

5.1 Controlling Mass Fuel Injection

Figure 4 shows ESCHER controlling the IMF to max-

imize the IMEP. At the beginning of the test, ES-

CHER does not know anything about the controlled

process. It ﬁrst decreases the IMF, but it is a mistake:

the critical level rises (ﬁrst lifecyle). The error is cor-

rected by increasing the IMF. The critical level drops,

so ESCHER keeps this action (lifecycles 2 and 3).

Then it perceives that while the IMF is rising, the crit-

ical level of the IMEP is rising too. But it is due to the

noise on the IMEP. This causes ESCHER to mistak-

enly decrease the IMF. The error is rapidly detected

and ESCHER eventually ﬁnds that it has to increase

the IMF in order to reduce the critical level (lifecycles

5 to 13). Once the maximal IMEP has been reached,

ESCHER oscillates around the IMF value that keeps

Figure 4: Controlling Injected Mass Fuel to Maximize

IMEP.

the critical level at its minimum. During this test, ﬁve

Contexts Agents were created and the optimum value

was reached in 12 lifecycles.

5.2 Controlling Mass Fuel Injection

and Ignition Advance

In this test, ESCHER controls the ignition advance in

addition to the IMF. Figure 5 shows the curves of this

test. Results are similar to the previous test, with ES-

CHER ﬁrst committing a mistake on both of the con-

trolled inputs, causing a rise of the critical level. Then

the error is corrected, before the controller stabilizes

the engine around the maximum IMEP.

The optimum IMEP was reached in 18 lifecycles.

During the 30 lifecycles of the test, ESCHER created

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

248

Figure 5: Controlling Injected Mass Fuel and Ignition Ad-

vance to Maximize IMEP.

19 Context Agents (10 were created by Mass Fuel In-

jected Controller Agent and 9 by Ignition Advance

Controller Agent).

The main challenge with the real engine compared

to generated black boxes is to deal with noise and la-

tency. As ESCHER quickly reacts to errors, the noise

is handled naturally. To overcome the latency be-

tween an action and its effects, ESCHER waits during

a parametrizable period of time between its actions

and its observations. For each test on the engine, a

lifecycle was executed every three seconds.

5.3 Discussion

ESCHER ﬁnds optimum values for all of the tested

operating points. As it is impacted by the noise on

the variables, the time taken to converge may vary.

The biggest differences were observed at 4000rpm

and 700mbar of pressure in the inlet manifold where it

took from 100 seconds up to 400 seconds to converge.

But it remains quicker than human experts doing it by

hand as it is currently the case in the industry. More

tests are planned, with more objectives and more si-

multaneously controlled variables.

6 CONCLUSIONS AND FUTURE

WORKS

In this paper, we presented an Adaptive Multi-Agent

System, called ESCHER, which controls heat en-

gines, basing its behavior on self-organizing agents.

We detailed its architecture and the agents interac-

tions.

ESCHER is based on the idea that the complete

understanding of how an engine works is not neces-

sary to its control. All the controller needs to know

is how the engine behaves. In other words our con-

troller is of black-box type: ESCHER only perceives

the engine’s inputs and outputs, but not its internal

mechanisms. The basic principle of our controller is

the following: the state of the engine inputs/outputs is

memorized when an action is applied and the conse-

quences of this action are recorded. This information

will be used to decide whether this action was good or

not in regard of the user-deﬁned desired engine state.

This means that the quality of the control improves

over time: at the begining the controller knows noth-

ing about the engine, but it perpetually learns from

its actions and quickly manages to control it. The

learning is parallel to the control, so ESCHER con-

tinuously self-adapts to the process.

No prerequisite knowledge other than the inten-

tions of the user (i.e. some criticality functions)

is needed by ESCHER. It is also able to satisfy

multi-criteria constraints on multiple inputs and out-

puts. Moreover, the independence between Controller

Agents gives a certain modularity to ESCHER. Each

Controller Agent (and its related Context Agents) is

a stand-alone MAS that can be plugged on a process

input. Regardless on what is controlling the other in-

puts, it will be able to synchronize its actions to per-

form a correct control.

We showed that ESCHER is able to deal with non-

linearity and multiple inputs and outputs, and that it

handles the noise and latency of a real engine, without

heavy knowledge about the engine. The only required

instantiation work is the translation of the user’s ob-

jectives into criticality functions.

By focusing on its learning skills, we consider to

Self-organizingAgentsforanAdaptiveControlofHeatEngines

249

make the ESCHER generic enough to be easily ap-

plied to all kind of systems. Multi-Agent Systems us-

ing a similar pattern have already been successfully

applied to the control of the temperature of a biopro-

cess (Videau et al., 2011) and are currently being ap-

plied in the contexts of ambiant systems (Guivarch

et al., 2012) and energy management of buildings.

Secondly, the behavior itself needs some improve-

ments. New mechanisms for a better resolution of the

problem of latency between the actions and their ef-

fects are required. Moreover, in large processes, some

inputs may not affect some outputs. The learning

of the process behavior could be enhanced by taking

into account this fact. Finally, the main limitation to

the application of ESCHER is the need for predeﬁned

criticality functions. Self-deﬁned criticality functions

could simplify make our controller even easier to in-

stantiate.

REFERENCES

Astrom, K. J. and Hagglund, T. (1995). PID Controllers:

Theory, Design, and Tuning. Instrument Society of

America, Research Triangle Park, NC, second edition.

Boes, J., Glize, P., and Migeon, F. (2013). Mimicking Com-

plexity: Automatic Generation of Models for the De-

velopment of Self-Adaptive Systems. In International

Conference on Simulation and Modeling Methodolo-

gies, Technologies and Applications (SIMULTECH),

Reykjavik. INSTICC Press.

Camazine, S., Deneubourg, J.-L., Franks, N. R., Sneyd,

J., Theraulaz, G., and Bonabeau, E. (2003). Self-

Organization in Biological Systems. Princeton Uni-

versity Pres.

Clair, G., Kaddoum, E., Gleizes, M.-P., and Picard, G.

(2008). Self-regulation in self-organising multi-agent

systems for adaptive and intelligent manufacturing

control. In Second IEEE International Conference on

Self-Adaptive and Self-Organizing Systems.

Colosimo, B. M. and Del Castillo, E., editors (2007).

Bayesian Process Monitoring, Control and Optimiza-

tion. Taylor and Francis, Hoboken, NJ.

Dabo, M., Langlois, N., Respondek, W., and Chafouk, H.

(2008). NCGPC with dynamic extension applied to a

Turbocharged Diesel Engine. In Proceedings of the

International Federation of Automatic Control 17th

World Congress, pages 12065–12070.

Di Marzo Serugendo, G., Gleizes, M.-P., and Karageorgos,

A., editors (2011). Self-organising Software - From

Natural to Artiﬁcial Adaptation. Natural Computing

Series. Springer.

Georg

e, J.-P., Gleizes, M.-P., and Camps, V. (2011). Co-

operation. In Di Marzo Serugendo, G., editor,

Self-organising Software, Natural Computing Series,

pages 7–32. Springer Berlin Heidelberg.

Guivarch, V., Camps, V., and Pninou, A. (2012). Context

awareness and adaptation in ambient systems by an

adaptive multi-agent approach. In International Joint

Conference on Ambient Intelligence, Italy.

Hagan, M. T., Demuth, H. B., and De Jesus, O. (2002). An

introduction to the use of neural networks in control

systems. International Journal of Robust and Nonlin-

ear Control, 12(11):959–985.

Kreisselmeier, G. and Anderson, B. (1986). Robust model

reference adaptive control. IEEE Transactions on Au-

tomatic Control, 31(2):127 – 133.

Lee, C. C. (1990). Fuzzy logic in control systems: Fuzzy

logic controller. IEEE Transactions on Systems, Man

and Cybernetics, 20(2):404–418.

Lee, J. H. and Ricker, N. L. (1993). Extended kalman ﬁlter

based nonlinear model predictive control. In Ameri-

can Control Conference, pages 1895–1899.

Montresor, A., Meling, H., and Babaolu, z. (2003). Mes-

sor: Load-balancing through a swarm of autonomous

agents. In Moro, G. and Koubarakis, M., editors,

Agents and Peer-to-Peer Computing, volume 2530 of

Lecture Notes in Computer Science, pages 125–137.

Springer Berlin Heidelberg.

Nikolaou, M. (2001). Model predictive controllers: A criti-

cal synthesis of theory and industrial needs. Advances

in Chemical Engineering, 26:131–204.

O’Hare, G. M. and Jennings, N. R. (1996). Foundations of

distributed artiﬁcial intelligence. Wiley-Interscience.

Soderstrom, T. and Stoica, P. (1988). System identiﬁcation.

Prentice-Hall, Inc., Upper Saddle River, NJ, USA.

Stengel, R. F. (1991). Intelligent failure-tolerant control.

IEEE Control Systems, 11(4):14–23.

Tyreus, B. D. and Luyben, W. L. (1992). Tuning PI con-

trollers for integrator/dead time processes. Industrial

& Engineering Chemistry Research, 31(11).

Videau, S., Bernon, C., Glize, P., and Uribelarrea, J.-L.

(2011). Controlling Bioprocesses using Cooperative

Self-organizing Agents. In Demazeau, Y., editor,

PAAMS, volume 88 of Advances in Intelligent and Soft

Computing, pages 141–150. Springer-Verlag.

Visioli, A. (2001). Tuning of PID controllers with fuzzy

logic. IEE Proceedings - Control Theory and Appli-

cations, 148(1):1–8.

Wang, H. (2001). Multi-agent co-ordination for the sec-

ondary voltage control in power-system contingen-

cies. Generation, Transmission and Distribution,

IEEE Proceedings, 148(1):61 –66.

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

250