AGENT-BASED SIMULATION OF ECOLOGICAL MODELS

AGENT-BASED SIMULATION OF ECOLOGICAL MODELS António Pereira Pedro Duarte CEMAS – UFP, University Fernando Pessoa Praça 9 de Abril, 349, 4249-004 Porto, Portugal E-mails: [email protected], [email protected]

KEYWORDS Ecological Modelling, Intelligent Agents, Simulation Models. ABSTRACT Coastal lagoons ecosystems are very complex to model and its global behaviour it’s hard to simulate. This complexity is even greater if the simulation model includes the intelligent entities like decision-makers involved in the system. Typical ecological simulations include only models of entities that have a very easy to predict behaviour (using mathematical equations and/or simple logical rules). However, in real ecological systems, man is heavily involved and his decisions are not based on fixed mathematical equations or simples rules but contrarily they are based on an in-depth analysis of the environment using his knowledge. This paper introduces an approach to intelligent ecological systems simulation using the agent-based approach. Intelligent entities are modelled as agents that have perception of their environment, reason using their knowledge and are able to change the simulated environment by using a given set of configurable actions. INTRODUCTION The management of coastal ecosystems is a very complex but essential part of any sustainable development strategy. Located between land and open sea, these ecosystems receive fresh water inputs, rich in organic and mineral nutrients derived from urban, agricultural and industrial effluents and domestic sewage, and are subject to strong anthropogenic pressures due to tourism and shellfish/fish farming. The interactions between coastal lagoons, land environment and sea boundaries reveal high physical, chemical and biological complexities, making the management decisions difficult to take and the consequences of these decisions very hard to predict. In Portugal, several state institutions are responsible for managing coastal ecosystems: ICN (Nature Conservation Institute), City Councils, Official Tourism Manager, the Navy and some more. When these ecosystems are integrated in a natural park, the Directorate of the Natural Park is also involved. The huge number of possible combinations generated by the different management decision options, the opposite

Proceedings 5th Workshop on Agent-Based Simulation H. Coelho, B. Espinasse, eds. (c) SCS Europe BVBA, 2004 ISBN 3-936150-31-1 (book) / 3-936150-32-X (CD)

Luís Paulo Reis NIAD&R - LIACC Faculty of Engineering - University of Porto Rua Dr. Roberto Frias, s/n 4200-465 Porto, Portugal E-mail: [email protected]

interests of some decision agents and stakeholders and the slowness of the decision process make it very difficult to implement efficient management policies. The fragmented legal framing of the administration responsibilities enables an overlap of abilities between two or more entities, leaving simple questions like: “Is it advisable to enlarge an opened navigation channel (due to tourism pressure)?” without an immediate answer. A positive answer to the previous question raises another complex one: “What will be the future consequences and/or benefits?” The ecosystem physical and biogeochemical complexities justify the construction of mathematical models to make predictions about their evolution. Generally, these models don’t take into account the influence of the management decisions taken by the responsible entities. This paper introduces a possible helper to fulfil this gap: an agent-based simulation of ecological models. The architecture for the simulation system described herein is based on the “intelligent agents” approach (Norvig and Russel 1995; Weiss 1999; Wooldridge 2002; Reis 2003). An agent may be defined as a self-contained software program, specialized in achieving a set of goals, by autonomously performing tasks on behalf of users or other agents. Agents are particularly useful in complex, inaccessible and dynamic environments as ecosystems or other biological systems. The study presented try to obtain answers to the question: “How to combine the use of mathematical models and Geographical Information System (GIS) with multiple decision makers, represented by different intelligent agents that collaborate with a Decision Support System (DSS), in the management of a coastal lagoon ecosystem (Ria Formosa, Algarve)?” The following sections describe the type of ecological modelling problems under analysis in this study with some examples, the Calibration Agent developed under this project to calibrate the generic ecological model used, the Agent-Based Simulation approach, its architecture and examples of the implemented agents and their interactions with the simulation. The paper concludes with some conclusions and pointers to future work. ECOLOGICAL MODELLING Ecological models are simplified views of nature used to solve scientific or management problems. These models

only contain the characteristic features that are essential in the context of the problem to be solved or described. Ecological models may be considered a synthesis of what is known about the ecosystem with reference to the considered problem, as opposed to a statistical analysis a model is able to translate our knowledge about the system processes, formulated in mathematical equations, and component relationships and not only relationships between data (Jørgensen and Bendoricchio 2001). Ecological models include physical, chemical and biological processes to describe the main features of the ecosystem in analysis. The description of the physical, chemical and biological processes and the system structure do not account for all the details. Carefully designed models, which include important processes and components, still omit details that aren’t important to the problem under consideration – many irrelevant details would cloud the main objectives of a model. However, these omitted details might have a strong influence on the predicted output those models produce (Jørgensen and Bendoricchio 2001).

The authors are particularly concerned with coastal lagoons and ecosystems modelling. These ecosystems are subject to strong anthropogenic pressures due to tourism and shellfish/fish farming. These factors are responsible for important ecosystem changes characterized by eutrophic conditions, algal blooms, oxygen depletion and hydrogen sulphide production.

The Sungo Bay is modelled as a 2D vertically integrated, coupled hydrodynamic-biogeochemical model. It is based on a finite difference bathymetric staggered grid (Vreugdenhil 1989) with 1120 cells (32 columns x 35 lines) and a spatial resolution of 500m (Figure 1). The model time step is 18 seconds. The model has a land and an ocean boundary. It is forced by tidal height at the sea boundary, light intensity, air temperature, wind speed, cloud cover and boundary conditions for some of the simulated state variables and solves the general 2D transport equation (Knauss 1997; Neves 1985). The hydrodynamic sub-model solves the speed components, whereas biogeochemical processes such as primary productivity and grazing, as well as physical processes such as sediment deposition and resuspension provide the sources and sinks terms. 16 km 122º25’E

37º10’N

37º00’N

122º35’E

17.5 km

In ecological models of aquatic systems, physical processes include flow and circulation patterns, mixing and dispersion of mass and heat, water temperature, settling of planktonic organisms and suspended matter, insulation and light penetration. The simulation of these processes is very important for setting up a good model of the whole ecosystem and detailed descriptions of them are available and widely accepted by modellers. One of the most important compromises is to find the optimal time and space scales of the model. Spatial grids acceptable for physical and chemical processes (10 to 100 metres) are very detailed for biological processes, and similarly, minutes or hours are good time scales for physical and chemical processes, but hours, days and months may be appropriate time scales for biotic components of an ecosystem (Jørgensen and Bendoricchio 2001). The space division must account for variations in horizontal and vertical dimensions. The simplest geometric representation is the zero-dimensional (0D) model, which simulates the system as a point and all changes are only time dependent. One-dimensional (1D) representation models assume that the system is characterized by a prevailing one-directional flow (horizontal or vertical) and the properties of the system vary along that direction and time. When the system is large enough to present sensible variation of the properties, vertical and/or horizontal division is required and two or threedimensional (2D or 3D) representations are more common. Models of deep large lakes, deep bays or large river estuaries are examples of these representations.

Examples of ecological models can be found in (Bacher et al. 1998; Hawkins et al. 2002; Duarte et al. 2003). Particularly complex is the last example: one ecological modelling developed for Sungo Bay, located in Shandong Province of People’s Republic of China.

Figure 1: Sungo Bay, including Model Domain and Bathimetry (m). Also shown, a part of the model grid with a spatial resolution of 500m (Duarte et al. 2003) CALIBRATION AGENT As in other fields of science, mathematical models used in the fields of ecology and environment science are based on a body of knowledge formed with not generally accepted theories, debatable or controversial hypothesis, questionable simplifications and a bundle of implicit or ambiguous assumptions, i.e., based on an imperfect understanding of the dynamics of the object systems. This leads to highly uncertain model results because of the uncertainty associated with model parameters and inputs and, sometimes, the uncertainty in model structure (Scholten and Van der Tol 1998). When an ecological model is built, those uncertainties are intrinsic to the model and the major problem is to quantify the quality of the simulations in order to recognize if a modification of the concepts, laws simulating the processes or model parameters would improve it (Mesplé et al. 1995). If the concepts and laws of the

simulated processes are well established, attention must be directed to deciding parameter values. Calibration of these parameters, i.e., defining appropriate values for each parameter in the simulation in order to approximate simulation results to reality, is a task of major importance and it’s the first task to do before the use of the model for simulation studies.

inter-variable relationships information (“Training relationships simulation” box). It stores the information in its “knowledge database” as a square matrix, synthesising the relationships between different classes.

Model calibration is performed by comparing observed with predicted data and is a crucial phase in the modelling process. It’s an iterative and interactive task in which, after each simulation, the “modeller” analyses the results and changes one or more equation parameters trying to tune the model. This “tuning” procedure requires a good understanding of the effect of different parameters over different variables. Evaluation of the result’s quality is an easy task with simple algorithms (ex. linear regression between predicted and observed data), the system can classify the results quality in a qualitative or quantitative scale. A more complex problem is the selection of new parameter values to use in the next iteration by the model equations, trying to maximize the model quality of fit. One way of doing this is to give to the software agent a list with all changeable equation parameters, all possible ranges for those parameters and let it exhaustively search through all available parameter combinations until it finds the optimal one. This is a very intensive computation process due to its uninformed (and thus not intelligent) search through the system’s tens or hundreds of equations and parameters. Research on this matter should therefore be focused on devising intelligent search techniques that may be able to use the modeller’s knowledge to guide the search. Knowledge about the behaviour of all system processes, possessed by the “modeller” in the traditional calibration processes, shall be used to guide the selection of the new values for the parameters contained in different mathematical relationships. During this study, an intelligent agent, learns this knowledge in three phases: - Building matrices that synthesize the interclass and inter-variable relationships; - Analysing the intra and interclass steady-state sensitivity of different variables to different parameters and among variables; - Iterative model execution, measuring model lack of fit, adequacy and reliability (Scholten and Van der Tol 1998) until a convergence criteria is matched. This methodology gives the Calibration Agent the generality needed to be able to calibrate “any” type of model. The calibration procedure diagram used by this agent is shown in Figure 2. The first step consists in choosing the model and checking if its database is populated. The agent runs the model for some time just to gather the interclass and

Figure 2: Calibration Agent Procedure Diagram The next step is used by the Calibration Agent to perform an exhaustive sensitivity analysis and to synthesize the results obtained in several matrices. First, the intraclass sensitivity is analysed (sensitivity of each variable to each parameter of its own class). Average sensitivities are calculated from parameter ranges defined by the user in the model database. The results of this analysis are stored in one matrix per class like the one exemplified in Table 1– class A with nA variables and mA parameters: changes imposed on parameter P1A affects variables V1A and V2A by 0.3 % and –0.6%, respectively. Table 1: Matrix Synthesizing Intra-Class Sensitivity Coefficients between Variables and Parameters Class A Parm P1A Parm P2A Var V1A 0.003 0 Var V2A -0.006 0.004 ... Var VnA 0 0

...

Parm PmA 0.34 0 -0.0071

Secondly, the inter-class sensitivity is analysed (sensitivity of each variable of each class is analysed with respect to all variables of each class by which it is influenced). During this step, the model runs (“Training sensitivity simulation” box) keeping all variables and parameters constant except those directly involved in sensitivity analysis. Since intra and interclass sensitivities may depend on the values of the parameters or variables with respect to which calculations are made, average sensitivities are calculated from variable and parameter ranges defined by the user in the model database. The results of this analysis are stored in one square matrix like the one

exemplified in Table 2 – class A, with nA variables, influences class z, with nz variables: imposed changes on variable V1A affect variable V2z by 0.3%; imposed changes on variable V2A affect variables V1z, V2z and Vnz by –0.1%, 0.05% and 0.0056%, respectively.

by the agents that influences the ecosystem is previously communicated to the model database; the simulation system runs and the results are presented to a decision support system to help managers decisions.

Table 2: Matrix Synthesizing Inter-Class Sensitivity Coefficients between Different Variables

-

0

0

0.0008

0 -4E-05

-

-

-

0.0001

-

-

-

Var Vnz

-

0 0.003 -0.001 0.0005

...

-

Var V2z

-

Var V1z

-

Var V1z -0.0001 8E-05 Var V2z 0.0035 6E-06 ... Var Vnz 0.0006 -3E-05

Var VnA

-

...

Var V2A

Var V1A Var V2A ... Var VnA

Class z

Var V1A

Class z

... Class A

Class A

0 6E-05

After the Calibration Agent acquired this basic knowledge about the simulated system, it starts the third step in the calibration process. At this stage, it selects the variable (Y) with the worst fit to observed data. Secondly, it selects the parameter (P) or variable (Z) of its own class or of another class to which it is more sensitive. If it is a variable (Z), it selects the parameter (Pz) to which this influencing variable is more sensitive. It changes the value of P or Pz in order to increase/decrease the average value of the variable under calibration (Y) towards the desired direction directly or indirectly through Z and runs the model. At the end of the simulation, it measures model lack of fit, adequacy and reliability, considering the overall model results and iteratively changes parameter values until the desired range of the influencing parameter has been completely covered. It then chooses the best value of P or Pz and proceeds to the next variable with the worst fit restarting the process. After the first round across all model variables under calibration, the process may be repeated until the desired fit is obtained, model improvement stabilized or the calibration is stopped by the user. AGENT-BASED SIMULATION SYSTEM After the calibration task concluded the model is ready for simulation studies and to produce results for helping management decisions. The architecture proposed here for the simulation system (Figure 3) is based on the “intelligent agents” approach (Norvig and Russel 1995; Weiss 1999; Wooldridge 2002; Reis 2003). Each intelligent entity involved with the ecological system is included in the simulation system as an agent (Natural Park Directorate Agent, Shell fish Farmer Agent, Official Tourism Manager Agent, City Council Agent, Navy Command Agent, …). Any action planned

Figure 3 - Agent-based Simulation System Architecture Agent Architecture The general model of each agent is very common (Iglesias et al. 1996). Its internal components are depicted in Figure 4. The Communication module is responsible for the interface with the environment (coding/decoding messages and network interface). The Knowledge Database module is where the agent stores the information about itself (self interests and services) and about what it knows from the environment (environment restrictions, other agents and applications, their addresses, services provided by them and their interests). The Control module is responsible for the policies of the communications (order and selection of the messages to process and queue messages to send) and actions (implementation of the different criteria regarding the attention to the external requests). For the Shellfish farmer Agents seed actions are regulated by the environment rules that allow seeding in some areas but not in the others. The “Actions Policy” element (in the Control module) must access the Knowledge Database module to know the allowed area to seed (“Environment Data” element) and the best quantity and type of bivalves to seed (“Self Data” element). When multiple agents work together, their behaviours must be coordinated in order to attain a global result. Coordination may be defined as the act of working together in a harmonious way to attain a common goal or agreement (Malone and Crowston 1991; Reis 2003). Sometimes agents can cooperate to reach a common goal. However, often agents don’t share one common objective and have opposite interests. In this case, there are conflicts to solve and a negotiation process must start to coordinate efficiently agents’ actions. In the simulations types used in this study, the existence of

agents with different interests in the ecosystem use (industry, fishery, tourism…) make necessary its coordination to provide an efficient shared use of the ecosystem, fulfilling the restrictions imposed legally.

sponsible for seeding and harvesting a specific type of bivalves. Based on its knowledge and its perception of the environment, it must decide the exact regions and times for seeding and harvesting each type of bivalve. Agents’ actions and perceptions consist on messages exchanged with the simulation application. The Shellfish Farmer Agent has three actions: seed bivalves, inspect the state of the bivalves growing and harvest them when it is appropriate. The definition of the concepts used in communication language follow the BNF notation (Naur 1960), inspired in Coach Unilang (Reis and Lau 2002; Reis 2003) and is represented in Figure 5. As an example the message message (5 Shellfish_Farmer Simulator seed (AreaNW1 now kelp (density 0.5 0.7) 15 ) )

Figure 4 - Agent Model Shellfish Farmer Agent An example of the type of agents we are concerned about is the Shellfish Farmer Agent. This agent is re-

was the fifth message sent by the Shellfish Farmer Agent to the Simulator application requiring to seed now 15 kg of kelps, with a density between 0.5 and 0.7, in the region AreaNW1.Perceptions for Shellfish Farmer Agents are the results for its actions – seed results, inspect results and harvest results. As an example the message:

::= message ( ) ::= [integer] ::= [string] ::= [string] ::= | | ::= | | ::= | | ::= define ( ) ::= seed ( ) ::= inspect ( ) ::= harvest ( ) ::= seed_result ( ) ::= inspect_result ( {( )}) ::= harvest_result ( ) ::= [integer] | now ::= ok | failed ::= [string] ::= | ::= | ::= ::= subtidal | intertidal ::= ( ) ::= sandy | sand_muddy | muddy ::= ::= ::= excellent | good | poor ::= land ::= {} ::= | (rect | quad | circle [real] | arc [real] [real] [real] [real] ::= (point [real] [real]) ::= [real] ::= scallop | kelp | oyster ::= | ::= (length [real] [real]) ::= (density [real] [real])

Figure 5 - Agent Messages Definition (BNF Notation)

message (47 Simulator seed_result (5 ok) )

Shellfish_Farmer

could be the answer to the previous message indicating that the seed action has been well succeed. CONCLUSIONS AND FUTURE WORK Agent-based simulation of ecological systems is a promising approach enabling to separate intelligent decisionmaking from biogeochemical simulation. Intelligent entities in the system are represented using the agent based approach and communicate with the main simulation module using an appropriate high-level communication protocol based on domain specific grammar. Using a calibration agent for model tuning enables to fully automate a very complex and tedious problem to solve manually. Experiences have been made to validate this approach for model calibration. More intensive tests are being performed with ecological models found in (Bacher et al. 1998; Hawkins et al. 2002; Duarte et al. 2003). The result of this work will be applied in the calibration of the Ria Formosa (Algarve) model, in the DITTY project (European Commission 2003) and in the intelligent simulation of this and other ecological systems. ACKNOWLEDGEMENTS This work was supported by the DITTY project - “Development of an Information Technology Tool for the Management of European Coastal Lagoons under the influence of river-basin runoff”, contract number EVK32001-00226, EU.

Jørgensen, S.E. & G. Bendoricchio. 2001. Fundamentals of Ecological Modelling, Elsevier Science Ltd, 3rd edition. Knauss, J.A. 1997. Introduction to Physical Oceanography, Prentice-Hall, Englewood Cliffs, NJ. Malone, T. & Crowston, K. 1991. Towards an Interdisciplinary Theory of Coordination, Technical Report 120, Cambridge, MA, MIT, Center for Coordination Science Mesplé, F., Troussellier, M., Casellas, C. & P. Legendre. 1995. Evaluation of simple statistical criteria to quality a simulation, Ecological Modelling 88 (1996): 9-18. Naur, P. (ed.). 1960. "Revised Report on the Algorithmic Language ALGOL 60.", Communications of the ACM, Vol. 3 No.5 (May), 299-314. Neves, R.J.J. 1985. Ètude expérimentale et modélisation mathématique des circulations transitoire et residuelle dans l’éstuaire du Sado, Ph.D thesis, Université de Liège. Norvig, P. & Russel, S.J. 1995. Artificial Intelligence: a modern approach, Englewood Cliffs, Prentice Hall. Reis, L.P. & Lau, N. 2002. "COACH UNILANG - A Standard Language for Coaching a (Robo) Soccer Team", in RoboCup 2001: Robot Soccer World Cup V, Eds. Andreas Birk, Silvia Coradeschi, Satoshi Tadokoro, Lecture Notes on Artificial Intelligence V.2377, 183-192, Springer. Reis, L.P. 2003. Coordenação em Sistemas Multi-Agente: Aplicações na gestão Universitária e no Futebol Robótico, PhD Thesis, Faculty of Engineering of the Univ. Porto. Scholten H. & M.W.M. Van der Tol. 1998. Quantitative validation of deterministic models: when is a model acceptable? Proceedings of the Summer Computer Simulation Conference, SCS, San Diego, CA, USA: 404-409 (ISBN # 1-56555-149-4) Vreugdenhil, C.B. 1989. Computational hydraulics, An introduction, Springer-Verlag Weiss, G. 1999. ed., Multi-Agent Systems: A Modern Approach to Distributed Artificial Intelligence, MIT Press. Wooldridge, M. 2002. An Introduction to Multi-Agent Systems, John Wiley & Sons, Ltd.

REFERENCES

AUTHOR BIOGRAPHIES

Bacher, C., Duarte, P., Ferreira, J.G., Héral, M. & O. Raillard. 1998. Assessment and comparison of the MarennesOléron Bay (France) and Carlingford Lough (Ireland) Carrying Capacity with ecosystem models. Aquatic Ecology 31 (4): 379 – 394. Duarte, P., Meneses, R., Hawkins, A.J.S., Zhu, M., Fang, J. & J. Grant. 2003. Mathematical modelling to assess the carrying capacity for multi-species culture within coastal waters, Ecological Modelling 168 (2003): 109-143. European Commission. 2003. “Development of an Information Technology Tool for the Management of European Southern Lagoons under the influence of river-basin runoff – The DITTY Project”. S.P.I.03.135, DirectorateGeneral of Joint Research Centre, Institute for Environment and Sustainability. Hawkins, A. J. S., Duarte, P., Fang, J. G., Pascoe, P. L., Zhang, J. H., Zhang, X. L. & M. Zhu. 2002. A functional simulation of responsive filter-feeding and growth in bivalve shellfish, configured and validated for the scallop Chlamys farreri during culture in China. Journal of Experimental Marine Biology and Ecology 281: 13-40. Iglesias, C.A., González, J.C. & Velasco, J.R. 1996. MIX: A General Purpose MultiAgent Architecture, in M.Wooldridge, J.P.Mueller and M.Tambe, editors, Intelligent Agents Volume II, Springer-Verlag.

ANTÓNIO PEREIRA was born in Porto, Portugal and obtained a degree in Electrotechnical Engineering and a MSc in Artificial Intelligence and Computation in the University of Porto. He worked a couple of years in industry (EFACEC – Systems and Electronics and Siemens). Since 2003 he is dedicated to research in the DITTY project and is a PhD student in Electrotechnical Engineering. His e-mail address is: [email protected]. PEDRO DUARTE was born in Lisbon, Portugal, obtained his degree in Biology at the University of Lisbon and his PhD in Environmental Sciences at the New University of Lisbon, in 1995. He has been involved in several European projects doing research in the field of Ecological Modelling applied to environmental problems such as carrying capacity estimation for aquaculture. His e-mail address is: [email protected]

LUIS PAULO REIS was born in Porto, Portugal and have a PhD in Electrotechnical Engineering (Coordination in Multi-Agent Systems) in the University of Porto. He has been researching in the area of (Multi-Agent) intelligent simulation for several years in different projects including FC Portugal simulated robotic soccer team – World and European champion in 2000. His email address is [email protected] and his web-page can be found at http://www.fe.up.pt/~lpreis.