AN ARTIFICIAL LIFE ENVIRONMENT TO CONTROL COMPLEX PROCESSES: AN INDUSTRIAL APPLICATION FOR ENERGY PRODUCTION PLANTS

AN ARTIFICIAL LIFE ENVIRONMENT TO CONTROL COMPLEX PROCESSES: AN INDUSTRIAL APPLICATION FOR ENERGY PRODUCTION PLANTS M. Annunziato1, I. Bertini1, M.Lucchetti2, A. Pannicelli2, S. Pizzuti1,2. 1ENEA, 2CS

contact: [email protected]

Abstract. This paper describes the use of evolutionary methods inspired by artificial life environments for the development of solutions connected to combustion problems based on the evolutionary properties. The idea proposed in this paper is aimed to developing a new approach to the optimisation and control of complex processes for energy production/consumption. This methodology is based on evolutionary optimisation and it started from some successful experiences in the dynamic characterisation (for diagnostics and control) for at least two industrial applications (oil field diagnostics and combustion dynamic characterisation). Furthermore an optimisation study has shown very interesting features of artificial life environments. The basic features of the proposed approach are: • dynamics based • no intensive modelling (progressive training directly from the measurements) • able to follow the plant evolution The essence of this approach could be synthesised by the following sentence: "not control rules but autonomous structures able to generate optimized-control rules". Good results have been obtained for dynamic analysis using the dynamic moments technique based on detection of attractor morphology. The results obtained for flame dynamics characterization are resulted better than the more classical nonlinear discriminants. The driving process is the dynamic building of a model on the basis of the observation of the effects that the regulation actions (acted by the operators or any other existing control systems) have on the plant performance. The powerful of optimization have been shown by artificial life: in the standard Traveling Salesman Problem the obtained results are surely comparable (better in some cases in respect to other algorithms). The proposed approach has been applying to a real scale waste incinerator in the context of the EU Project : “Development of Evolutionary Control Technology for Sustainable Thermal Processes” (ECOTHERM). The preliminary results are described.

1 2

Italian Agency for New Technology, Energy and Environment, Roma – ITALY Communication &Systems, Roma - ITALY

Introduction

The ideas of evolution, complexity, intelligence and life reproduction have long been stimulating the collective thinking. Scientific approaches then become predominant on the formation of hypothesis and practices to answer to these basic questions. Research and development, inspired by mathematical and physical models of intelligence (Artificial Intelligence) and more recently of life itself (Artificial Life), are providing new tools and ideas for the solution of complex problems requiring evolving structures. In problems ranging from traffic regulation to energy process control and optimisation the not-adaptive approaches are not effective to solve the problem over the time. The uncontrolled variables, the process ageing, the unforeseeable effects caused by human errors, the evolution of the process, in most of the cases require the change of the basic model or the objectives, or even the whole strategy. To reach the goal of evolving structures, a continuous learning of the system from the environment is necessary but not sufficient, and the ability of the system to change its internal structure is needed. In short, we need information structures able to evolve in parallel to the process we are modelling. Since late 70's a new branch of theory has been introduced in the evolutionary system research: the genetic algorithms, starting from [16] and developed in different directions [14]. In these approaches the algorithm structure is able to optimise a fitness function, or to optimise a winning strategy simulating some mechanisms of the genetic dynamics of chromosomes (reproduction, recombination, mutation, selection). These algorithms have been successfully applied in many technological and engineering problems, in order to solve optimisation [18] or design problems. The limitation of these approaches is that the internal structure of the information is generally static and defined/controlled by the author of the algorithm. The fusion of the concepts of genetics algorithms and the self-organisation brought about the concept of the artificial life [9][12][17] started in the 80's. For the first time, it has really opened the possibility to build evolving structures able to develop a completely new organisation of the internal information. Artificial life is generally applied to study biological and social systems using software simulators [17][19] and the basic concept is to leave the system with the necessary degree of freedom to develop an emergent behaviour, combining the genetics with other life aspects (interaction, competition, co-operation, food network, etc.). At present, artificial life (or alife) is used mainly to study evolution problems, but we think that it has the potential to generate information structures that are able to develop a local intelligence. With the term local intelligence we refer to an intelligence strongly connected to the environmental context (the problem) we need to solve. We are involved in the development of this kind of structures called artificial societies [9][13]. The goal of these structures is the solution of a specific class of complex problems (design and engineering [1][2][3]) which require evolving structures. The extensive use of energy presents a severe challenge to the environment and makes indispensable to focus the research on the maximization of the energy efficiency and minimization of environmental impact (in particular the reduction of NOx and CO emissions). In this context the combustion process control assumes an importance much more relevant with respect to the past, especially for the combustion plants where the pollutants emissions, the environmental impact and the energy efficiency are strictly related to the modality of the process management. The proposed methodology is based on dynamics-based classification and evolutionary optimization. The principal features of the approach are: dynamics based, no intensive modeling (progressive training directly from the measurements), ability to follow the process evolution. In our proposal the process knowledge is developed directly by the system through the observation of the effects that the regulation actions (acted by the

operators or any other existing control systems) have on the plant performance. The main processes which we are looking at for application of the evolutionary control in the context of combustion plants are: eco-sustainable energy processes, gas turbines, industrial combustion chambers, engines. Actually, we are using this methodology in order to develop a prototypal control system for a real waste incinerator plant.

The evolutionary control In complex processes, one of the basic problem we have to face for control is the continuous evolution of the plant along the life (aging, maintenance, upgrading, etc..). This is a big problem for traditional control methods: being based on fixed optimization rules, they don’t take care of the evolution of the process during its life. In order to try to overpass this difficulty evolutionary computation techniques have already been applied to non stationary environments [10][15][21] and we developed an evolutionary adaptive technique, named evolutionary control, oriented to the optimization and control of complex systems in non stationary environments. The basic features of the methodology we propose are: • no intensive modeling (progressive training and updating directly from measurements); • capability to follow the process evolution. The essence could be synthesized by the sentence: "not control rules but autonomous structures able to dynamically generate optimized-control rules”. In our proposal, the process knowledge is obtained directly by the system through measurements observation, and it’s used to update a dynamic model of the process itself, which we call performance model. The implementation of this idea is described in figure 1. The basic concept consists in the realization of an artificial environment that lives in parallel with the process and that asynchronously communicate with it, in order to dynamically control and optimize it. We suppose to always measure from the process its current regulations and the related value of an observable quantity which we call performance and which represents the objective function of our optimization. In this way measurements are composed by both process variables and performance. The system continuously gets measurements and the dynamic state description from the process and provides the process back with the control actions. The main blocks of the evolutionary control (fig. 1) are the alife environment and the performance model. Performance

Model

Evolutionary Control

PROCESS

measurements new Control

control actions

best

Artificial Life Environment (ALIFE)

Fig. 1. Scheme of the evolutionary control The first one is an artificial environment composed by individuals able to find the optimal solutions. The second one is a model of the process performance used by the

alife environment in order to provide its individuals with a fitness value. The final control actions are the average between the best solution achieved by the alife environment and the current regulations. This is because we wish smooth transitions among different states. Each time a new measurement is acquired the performance model is updated with it (continuous learning) and a new individual, representing the new experimented/observed process condition, is inserted in the artificial environment. Thus the system is continuously updated, it follows the process notmonitored changes and drives the evolution towards better performances. Of course, at the beginning the system is not able to give any suggestion but it only learns from the process measurements. The artificial environment starts being active and giving right suggestions when the performance model is trained.

The artificial life environment for on-line optimization The artificial environment implemented derives from the Artificial Society approach [2]. This approach has been tested for the optimization of a static well known problem, the Traveling Salesman Problem, where it has reached the optimal value for the 30, 50 and 75 towns [2]. The alife context is a two-dimensional lattice (life space) representing a flat physical space where the artificial individuals (or autonomous agents) can move around. Every iteration (life cycle), the individual moves in the life space and, in case of meeting with other individuals, interaction occurs. Each agent has a particular set of rules that determines its interactions with other agents basically based on a competition for energy in relation to the performance value. Agents can also self-reproduce via haploid mutation. We suppose to periodically acquire a set of measurements on the real process (measurement cycle), to calculate the current value of the actual performance and to provide such information to the control system. The performance is the target of the optimization and it is derived from measurements. At every cycle of measurement, a new individual is built on the base of the measured values and inserted in the environment. Three blocks compose the data structure of each individual: the genotype, the information and the status. The first one includes a collection of behavioural parameters regarding dynamics, reproduction and interaction. The information block includes a series of parameters related to the process to control: the regulation and measurement values; both the information and the genotype don't change during the individual life. The status parameters include dynamics and structural parameters (position, direction, curvature, wire description), age, energy and performance values. These parameters change during the individual life. An important feature that makes us think of this strategy as the right one is the finiteness of the individual life. The optimization, in fact, succeeds in keeping itself updated on the evolution of the process by continuously renewing the population on line. This is allowed by an ageing mechanism, according to which each individual dies after a fixed number of life cycles (average life). The performance is updated using an external problemspecific model. This is due to the possible changes in the unknown variables of the process not represented in the genotype. For this reason the performance variable is located in the status block. We can summarize the main issues of alife approach as follows: • biodiversity – the periodic introduction of new individuals lets the algorithm not to converge towards a local optimum, but several search path remain active in order to allow a deeper exploration of the regulations space;

•

evolution – the artificial environment is able to evolve in parallel with the process, continuously renewing the whole configurations among which the proposed optimal control set is chosen; • adaptivity – owing to the above-mentioned evolution and biodiversity, the system is able to quickly react whenever an unforeseen change in fitness landscape occurs. A detailed description of the method and artificial life environment is reported in [3].

Fig. 2. The Artificial Life environment

The performance model The task of the performance model is to provide the newly generated individuals of the alife environment with a performance. The performance evaluator has in input the values of the control parameters and returns an estimate of the corresponding value of the fitness function. At present this module is implemented with a performance map. With this implementation the control parameters are discretized and the values we obtain are intended to be the performance table indexes. In the figure we show the situation with two control parameters indexing the X and Y axis of the table. We have two possible cases, depending on this discretization of the input parameters (fig. 3): 1. If the input refers to a cell in which we have already inserted a measure then we will consider that quantity as the estimate of the performance we are looking for; 2. If the input refers to an empty cell we use an interpolation mechanism, according to which if we point to an empty cell, we have to search for an already measured value in the neighbourhood, within a fixed range. In particular we stated as interpolation rule the following one.

measurement Regulation s interpolation

Fig. 3. Performance evaluation

In particular we chose a linear interpolation rule K (d )⋅ M f + (1 − K (d ))⋅ R

(1)

where Mj is the measure found in the neighbourhood of the empty cell, R is a random performance estimation of the point and K(d) is a weight coefficient exponentially decreasing with the distance between Mj and R in the parameters space. The performance map has a twofold task. On one hand it is, as we already said, the long-term memory of the control system; on the other hand, it allows the continuous updating of the reference model and so it lets the control system itself to evolve in parallel with the process. So it roughly represents an internal knowledge of the real system. The process of updating of the performance map is rather simple. When there is a new measurement, as we said above, this is inserted in the table in the cell which the discretization of the parameter set refers to. This happens regardless of the previous filling of that particular cell. However this simple knowledge implementation has several drawbacks. First the discretization of the map severely affects the performance of the system. Second it only allows of low dimensional systems, systems with a few control parameters which the performance depends on, because of memory allocation. Third the interpolation mechanism assumes the performance behaviour to be locally linear. In order to overcame these drawbacks a performance module based on evolutionary neural networks [4][6][7] is being carried out to replace the performance map.

Applications As a first step, the evolutionary control has been tested on the KuramotoShrivashinsky [20] differential equation system which describes the propagation of unstable flame front in uniform combustible mixtures. Experimentation [2] (fig. 4) concerned to maintain at 1 m/sec the flame propagation velocity. The effect of the control reduce the deviations induced by artificial disturbances inserted on the system.

1.6

Controlled Not controlled

1.5 1.4

velocity

1.3 1.2 1.1 1 0.9 0.8 0.7 97

91

85

79

73

67

61

55

49

43

37

31

25

19

13

7

1

0.6

time iterations

Fig. 4. Comparison between a controlled and a not-controlled conditions in the KuramotoShrivashinsky model.

Subsequently the method has been tested on the well known chaotic Chua’s electronic circuit [11] (in software and hardware), the experimental bench we designed and implemented is sketched in Figure 5, where it is shown a block diagram of the hardware/software tester device. It is composed by the hardware of the Chua’s circuit, by the device for parameters’ regulation, by the PC and by the software for analysis and control.

Fig. 5: experimental bench The blocks belong to different hardware: PC and an external device. Blocks 2, 3, 4 and 7 are “software blocks”: block 2 and 3 include information to estimate the performance function, block 4 is the control algorithm and block 7 is the software for analysis, control and management of the whole device. Blocks 1 and 5, which have been designed and built in ENEA, represent the chaotic oscillator, based on the Chua’s model, with relative electronic accessory and the electronic device for the communications between PC and circuit (Figure 6). The operation of the whole device is synthesized as follows: the “performance calculation” (block 2) estimates the signal produced by the Chua’s circuit (block 1) to identify the level of difference of this signal with the one chosen as target (the elements that characterized target signal are in block 3); this operation gives rise to a performance index, through which the Alife algorithm (block 4) produces a new set of parameter values for the Chua’s

circuit; in the following iterations, the signal dynamics will direct to identify itself with the wanted one.

Fig. 6: oscillator In all these experiences we use some (two or three) parameters to control the system and an external parameter as a unknown noise continuously varying. Finally we leave the control system to manage the regulation parameters in order to continuously optimize a performance function. In [2][3] detailed results about the on-line optimization capability of the system are reported. p e r f o r m a n c e

1

evolutionary control

0.8

0.6

0.4

without control

0.2

0 0

200

400

600

800

1000

1200

1400

1600

1800

Fig. 7. Typical result of the performance recovery of the evolutionary control. As example, fig. 7 shows the results of the control on the Chua’s system that is characterized by 4 control parameters. In this experiment we try to regulate three parameters in order to recovery the performance of the circuit which is continuously disturbed by external changes on the fourth parameters. The curve of the performance (blue) obtained with the evolutionary control is compared with the curve of the performance (dotted red) in absence of any sort of control. In this experimentation the average performance for the uncontrolled situation is 0.5033 while in the controlled case it is 0.842 achieving a 34% improvement on the average performance.

Another benchmark concerns the traffic prediction. The use of neural computing for transportation applications began recently and work has largely been of an exploratory nature. Applications which have been addressed using artificial neural networks (ANN) range from forecasting/classification of traffic flow parameters/traffic states to incident detection, from driver behaviour/vehicle control to traffic control and traffic monitoring. In nearly all of the applications reviewed the back-propagation learning algorithm was used. Despite some encouraging results its main drawback is the lack of on-line adaptation to changing conditions. For artificial neural networks to be viable for on-line applications in transportation they will need to be able to function in real time. Recently another interesting area is the one concerning the application of evolutionary computation based methodologies to traffic control, traffic management, traffic signal operation. What we propose is an innovative approach for traffic prediction which combines these two methodologies in order to carry out evolutionary neural models capable to on-line and dynamically adapt to changing conditions giving rise to a new class of ANN called evolutionary neural networks (ENN). Experimentation concerned the 20 minutes prediction of traffic flow rates using neural networks. The goal is to optimise the weights of a neural network structured with 8 input nodes, three hidden nodes and one output node (the traffic flow rate forecast) using the standard Back-Propagation algorithm (BP) and the above mentioned evolutionary algorithms in order to compare off-line and on-line approaches. In both situations that we used as transfer functions for all the nodes the classic sigmoid (2). -x

Y = 1/(1+e )

(2)

The data set consists of one week observations, 1980 data as result of a 5 minutes moving average filter, of the vehicles flow rate of the Genoa’s urban freeway downloaded from “The EUNITE SAS data repository”. Training and testing sets are exactly the same for all experimentations and the whole data set has been partitioned into 1800 training data and 180 testing data. In the first bunch of experiments a direct comparison with the BP algorithm can be done, in the second one the network is optimised on a travelling window of the last ten data (50 min.). In this situation every time the data set changes different weights are dynamically found, in this way the neural model is capable to adapt in real-time to changes. Table 1 and 2 show a comparison of the main experimental training and testing results. Results are calculated according to the RMSE formula (3). RMSE =

1 n

n

∑ (0.5 * ∑ 1

(3)

m

( y − yt ) 2 )

1

In tables 3 and 4 the main experimental settings for off-line and on-line optimisation are shown.

Off-line On-line

BP 0.1

CP 0.043 0.013

Table 1: training results

PE 0.038 0.005

Off-line On-line

BP 0.11

CP 0.042 0.031

PE 0.035 0.033

Table 2: testing results

BP CP PE

Generations/Cycles 3000000 500 2000

120Max population size 100 256

Table 3: Experimental settings used for off-line training

CP PE

Generations 50 100

Max population size 100 256

Table 4: Experimental settings used for on-line training Tables 1 and 2 show very encouraging results. In the off-line situation evolutionary methods show a remarkable increase of accuracy performance compared to the BP algorithm. The on-line experimentation shows a significant improvement with respect to the previous case as well. The reason for this is pretty simple because in the off-line situation a difficult overall global model is built up while in the on-line case several easy local models are dynamically made. These results clearly show the effectiveness of using evolutionary methodologies to build up adaptive neural models overcoming the off-line drawbacks imposed by the BP based methodologies. Finally the experimental qualification on large scale plant have been started in the frame of the EcoTherm 5FP-EU project (2002-2004: “Evolutionary Control for Thermal sustainable processes”). Two plants for thermo-valorization of solid urban waste are considered (Ferrara plant in Italy and Rotterdam plant in Netherlands). Preliminary results are reported in next paragraph. The application to urban waste thermo-valorization plant The proposed approach is being applied at a real scale waste incinerator of a multiservices special company of Ferrara Municipality (Italy), the Azienda Gas Energia Ambiente (AGEA), which has as principal scope the management of energyenvironmental services on the territory. The plant of Canal Grande (located close to the city) in connection with a geothermal power plant provides the heating of part of the civil habitations and produces the 37.5% of the annual requirements of the users. The plant, designed in 1988, is working since the end of 1993 and can be considered a modern plant. It actually respects all the EU directives and imposed limits. The technology used for the thermal destruction is classified as grid furnace, which has a wide use in the waste combustion area, particularly three steps grid with alternate mechanic movement. The combustion chamber is characterized by the following parameter: feed flow rate 400000-800000 Kcal/(m2*h), specific mechanic charge

200-400 Kg/(m2*h), specific thermal charge 60000-200000 Kcal/(m3*h). Due to the previous Italian regulation the plant is provided with post-combustion chamber in order to guarantee a controlled permanence of the gas produced by the combustion before the out of the chimney. The plant can treat 40000 tons/year of urban solid waste, which have a PCI average of 2500 Kcal/kg and a capacity of 6tons/hour. See fig. 8 for the plant layout. The plant is monitored by several different equipment: the Flame Dynamic Detection system in order to identify the flame dynamics through “dynamic moment” analysis [2] of the of the signals obtained by image sequences acquired inside of the combustion chambers; the real-time data acquisition system in order to acquire all regulations variable of the plant; a chemical species commercial analyser and chemical species predictor (developed by ENEA) in order to measure the plant performance.

Fig. 8. The AGEA-Ferrara waste incinerator

Fuzzy performance definition The measured performance model is aimed to provide the evolutionary controller with a global index of the performance which takes into account all the variables and constraints. The definition of this index is problem specific, multi-objective approach and it is made with help of the process engineers. It is represent the global performance of the system we want to maximize. In order to carry out the performer measurement it has been chosen to use fuzzy sets theory to properly define and compose the different variables and criteria. The fuzzy approach has been also chosen because it allows the operators transparency, it provides a well established theoretical framework to solve this kind of problems (providing a global index in the lattice [0,1]) and it is highly flexible because it can be transported to different plants with little effort. First the basic fuzzy sets are introduced (membership functions definition), then two different composition criteria are described and finally performance is defined. In table 5 fuzzy sets and membership functions are described.

Fuzzy set “Average steam flow rate ‘good’ ” “Average Steam flow rate ‘stable’ ” “Average O2 ‘good’ ” “Average temperature ‘good’ ” “Average NOx emissions ‘low’ ” “Average CO emissions ‘low’ ” “Average flue gas rate ‘low’ ” “Average waste flow rate ‘high’ ”

Membership function Trapezium Gaussian Triangle trapezium trapezium trapezium trapezium ramp

Table 5. fuzzy sets definition

The main idea driving the definition of the fitness criterion is that of having a flexible function capable to manage different criteria. In particular the fitness function will be the composition of two fuzzy sets describing two different requirements : ‘optimality’ and ‘strictness’. The difference between the two lies in the composition of the previously defined fuzzy sets. The membership function of the first one will be defined as the weighted sum of the membership functions of basic fuzzy sets. Logically this operator represents a composition standing between AND and OR. This fuzzy set will fulfil the ‘optimality’ requirement because it allows to set the weights, the importance of each fuzzy set, according to the custom needs. In this way the optimiser will find the optimal solution for that particular setting of the weights giving the system scalability to different needs. The second criterion will concern the strict constraints satisfaction defined in the basic fuzzy sets. The resulting fuzzy set will be logically defined as the AND composition of the basic fuzzy sets. It means that the resulting membership function will be the product or the minimum of the membership p e rfo rm a n c e 1

0 .9

0 .8

memb.value

0 .7

0 .6 p e rf o p t. c o n s tr a in ts fitn e s s

0 .5

0 .4

0 .3

0 .2

0 .1

4248

4111

3974

3837

3700

3563

3426

3289

3152

3015

2878

2741

2604

2467

2330

2193

2056

1919

1782

1645

1508

1371

1234

960

1097

823

686

549

412

275

1

138

0

ti m e

Fig. 9. Fuzzy performance index behaviour functions of the basic fuzzy sets. The final fuzzy set describing the global fitness will be the weighted sum of the last two fuzzy sets. Weights will be defined by the developer according to the custom requirements depending on the relative importance

of the two criteria. In figure 9 it is shown as example how the fuzzy performance index behaves on real data. The reader interested in further details can refer to [5].

Conclusion A new approach for the control and on-line optimization has been described in its main components. This approach is based on dynamic state identification of the system and evolutionary optimization on the process regulations. Good results have been obtained for dynamic analysis using the dynamic moment technique based on detection of attractor morphology. The results obtained for flame dynamics characterization are resulted better than the more classical non-linear discriminants. The powerful of optimization have been shown by artificial life. A scheme of the overall implementation of the control strategy has been described. Software control simulators (Kuramoto-Sivashinsky equation and Chua electronic circuit) have been utilized to point out the control ability of the whole architecture. Finally it has started the application of the proposed strategy on real scale waste incinerator in the framework of the EU Ecotherm project. First results show very encouraging confirmation about usefulness of dynamic description in order to improve the modeling of plant performances. More detailed studies are in progress to evaluate the performances of the control methodology in terms of a) which is the response time of intervention of the system, b) which disturbances can be recovered, c) which kind of new optimized configurations can be generated.

REFERENCES [1] Annunziato M., Bertini I., Piacentini M., Pannicelli A. , 1999, “Flame dynamics characterisation by chaotic analysis of image sequences” 36 Int. HTMF Institute, Sacramento/CA, USA [2] Annunziato M. , Bertini I. , Pannicelli A., Pizzuti S. , Tsimring L. , 2000 , “Complexity and Control of Combustion Processes in Industry”, proceedings of CCSI2000 Complexity and Complex System in Industry, Warwick, UK [3] Annunziato, M., Bertini, I., Lucchetti, M., Pannicelli, A., Pizzuti, S. , 2001 , Adaptivity of Artificial Life Environment for On-Line Optimization of Evolving Dynamical Systems, in Proc. EUNITE01, Tenerife (Spain). [4] Annunziato M., Bertini I., Lucchetti M., Pizzuti S., 2003, “Evolving Weights and Transfer Functions in Feed Forward Neural Networks” , in Proc. EUNITE03, Oulu, Finland [5] Annunziato M., Bertini I., Lucchetti M., Pizzuti S., van Kessel L.B.M. , Arendsen A.R.J., 2003, AN EVOLUTIONARY ADAPTIVE MODEL FOR A MSWI PLANT, Deliverable D4 of Task 1.4: Evolutive process model, EU Project ECOTHERM [6] Annunziato, M., Bertini, I., Pannicelli, A., Pizzuti, S. , 2003, “Evolutionary feed-forward neural networks for traffic prediction” , proceedings of EUROGEN2003, Barcelona, Spain [7] Annunziato M., Lucchetti M., Pizzuti S., 2002, “Adaptive Systems and Evolutionary Neural Networks : a Survey”, in Proc. EUNITE02, Albufeira, Portugal. [8] Annunziato M. , Pizzuti S. , "Fuzzy fusion between fluidodynamic and neural models for monitoring multiphase flows", International Journal of Approximate Reasoning, vol.21, issue 3, 249-267, 1999

[9] Annunziato, M., , “Emerging Structures in Artificial Societies”, in Creative Application Lab CDROM, Siggraph, Los Angeles/CA, USA, 1999 [10] Branke J. , 1999, “Evolutionary Approaches to Dynamic Optimisation Problems – A Survey”, GECCO Workshop on Evolutionary Algorithms for Dynamic Optimization Problems, A. Wu (ed.), 134-137 [11] Chua L.O., 1992, "The Genesis of Chua's Circuit", AEU 46, 250 [12] Emmeche, 1991, “The Garden in the Machine : The Emerging Science of Artificial Life”, Princeton University Press. [13] Epstein J.M. and Axtell R.L., 1996, “ Growing Artificial Societies : Social Science from theBottom Up (Complex Adaptive Systems)”, MIT Press, Cambridge/MA, USA [14] Goldberg D. E. , 1989, “Genetic Algorithms in Search, Optimisation and Machine Learning”, Addison Wesley, USA [15] Grefenstette J. J. , 1992, “Genetic algorithms for changing environments”, R. Maenner and B. Manderick, editors, Parallel Problem Solving from Nature 2, North Holland, 137144. [16] Holland J. H. , 1975, “Adaption in Natural and Artificial Systems”, MIT Press, Cambridge/MA, USA [17] Langton. C. , 1989, “Artificial Life”, C. Langton Ed. Addison-Wesley. pp. 1-47. [18] Oliver and Smith D. and Holland J. R. , 1987, “A study of permutation crossover nd operators on the travelling salesman problem”, Proc. of the 2 International Conference on Genetic Algorithms, J.J. Grefenstette ed., Lawrence Erlbaum, Hillsdale/NJ, USA, 224230. [19] Rocha L.M., “Evolutionary Systems and Artificial Life”, Lecture Notes, Los Alamos National Laboratory, 1997. [20] Sivashinsky G.I., Nonlinear Analysis of Hydrodynamic Instability in Laminar Flames Acta Astronomica, 4, 1177, 1977. [21] Trojanowski K. and Michalewicz Z. , 1999, “Evolutionary Algorithms for Non-Stationary Environments'”, Proceedings of 8th Workshop: Intelligent Information systems, Ustron, Poland, ICS PAS Press, pp 229-240.