Developement of Multi-Agent system (MAS) model for Bac Lieu case study
The objective of the past work was to capture the main processes, both
biophysical and economic, to reproduce them in a virtual world similar to reality. There have been two stages. In a first step, the team tried to select relevant knowledge to set-up a first conceptual model. As an hydrological model was already developed (Hoanh, et al. 2001), the emphasis was put on the economic and social processes. The first idea was to work on the interactions between decisions making at different scales. Are the decisions making processes of the different stakeholders (farmers, traders, hamlet, commune, water company, etc.) well......
Development of Multi-Agent Systems (MAS) model
for Bac Lieu case study
--- oOo ---
F. Bousquet , LC. Dung2, LA. Tuan2
1
1
IRRI-Cirad, Bangkok, Thailand
2
Cantho University, Vietnam
1 INTRODUCTION
Beside the DFID project “Accelerating poverty elimination through sustainable
resource management in coastal lands protected from salinity intrusion” a
complementary modeling experiment started in September 2001. The
objective is to test the use of a new methodology using Multi-Agent Systems
for participatory modeling. This approach was initiated by Cirad (Bousquet, et
al. 1999) and has already been developed in collaboration with IRRI. For
instance this approach is developed and used by IRRI researchers in North
Vietnam for the study of farmer’s practices (Castella, et al. 2001).
MAS is a methodology which put emphasis on the understanding of decision
making processes of different stakeholders, linked to bio-physical dynamics.
Thus, it seems to be a method complementary to the others method already
used in DFID project (spatial analysis for land management unit, hydrological
modeling and economic survey). It may be helpful for the integration of
knowledge coming from different disciplines.
It has been decided to conduct a one year project on MAS modeling to train
the participants and assess if this method is relevant and of some interest for
the case study. The team is composed of a modeler and two scientists from
Cantho University specialized in Economics and Hydrology. The IRRI-DFID
project experts are also participating by delivering the expert knowledge and
giving access to the data collected. In January 2002 the team met for the first
time with the DFID project to discuss on the issues and dynamics at stake.
Then, during April 2002 the modeling team produced a conceptual model of
farmers’ decision-making process that is presented in this paper. The
objective is to inform and provoke discussions with researchers and
stakeholders and orient the research for the remaining six months.
The objective of the past work was to capture the main processes, both
biophysical and economic, to reproduce them in a virtual world similar to
reality. There have been two stages. In a first step, the team tried to select
relevant knowledge to set-up a first conceptual model. As an hydrological
model was already developed (Hoanh, et al. 2001), the emphasis was put on
the economic and social processes. The first idea was to work on the
interactions between decisions making at different scales. Are the decisions
making processes of the different stakeholders (farmers, traders, hamlet,
commune, water company, etc.) well coordinated? What is the influence of a
stakeholder’s decision on another one?
1
Then, a research question has been proposed in April 2002 by the
researchers from University of Cantho: how rice/shrimp systems affect
economic differentiation? The associated hypothesis is that rich farmers will
get richer by applying shrimp production system, poor will get poorer. The rice
farmers will get poorer because of low rice price, and as there is high risk of
shrimp failure the poor may have difficulties to recover from these failures.
In this paper, we present firstly some principles of multi-agent systems. Then
we present the model, which have been designed and partly implemented. It
has been implemented with the Cormas platform (Bousquet, et al. 1998). The
design of the model leads to new questions. Lastly we propose a discussion
on the next steps and possible perspectives for this research.
2 PRINCIPLES OF MAS
2.1 What is a MAS?
The aim of multi-agent systems (Ferber 1999, Weiss 1999) is to understand
how independent processes in direct competition are coordinated. An agent is
thus a computerized process, something that comes between a computer
program and a robot. An agent can be described as autonomous because it
has the capacity to adapt when its environment changes. A MAS is made up
of a set of computer processes that occur at the same time, that is, several
agents exist at the same time, share common resources and communicate
with each other. The key issue in the theory of MAS is formalizing the
necessary coordination among agents. The theory of agents is therefore a
theory of
1. Decision-making: what decision-making mechanisms are available to the
agent? What are the linkages among their perceptions, representations,
and actions?
2. Control: What are the hierarchical relationships among agents? How are
they synchronized?
3. Communication: What kinds of messages do they send to each other?
What syntax do these messages obey?
For which elaborate formulas are put forward.
Here are the key questions that can be explored using MAS: How do
individuals make up a group? How is an institution created? The individual
cannot be considered as an autonomous entity that is independent of its
social environment. Therefore, how are individuals constrained by collective
structures that they themselves have set up and how do they make these
structures evolve (Gilbert 1995)? What degrees of freedom are given to the
definition of individual practices? Here are just some of the questions that can
be explored using MAS and that can be summarized as follows: How are
collective structures set up and how do they function when they are based on
agents with different capacities of representation, that exchange information,
goods, or services, etc., draw up contracts, and are thrust into a dynamic
environment that responds to their actions?
2
We think that such questions are relevant to some important issues in INRM,
and thus the interdisciplinary research was very promising when some of us
started to work on this methodology ten years ago. This was confirmed by the
successful development of several applications (Barreteau and Bousquet
2000, Bousquet, et al. 2001, Rouchier, et al. 2001).
2.2 The use of MAS
The new INRM approaches developed by the CGIAR go with new scientific
opportunities and breakthroughs, among which spatial modeling is a key one.
“Models will have increased capacity to integrate social and bio-economic
information”(Bilderberg 1999)
Adaptive management is defined as “building and maintaining stocks of
biodiversity that ensure that functional integrity of the system can increase the
adaptive capacity. Adaptive capacity is dependent on knowledge — its
generation and free interchange — the ability to recognize points of
intervention and to construct a bank of options for resource management”.
The role of modeling is formulated in this context: “Modeling proceeds
iteratively by successive approximations usually from simple to more complex
representations of system dynamics. This iterative modeling is done in close
interaction with stakeholders, who, along with the modelers, use the models
for scenario planning.”
This is how we would like to use the models: we try to integrate knowledge
from various sources in a iterative way. Then the model is used with the
stakeholders to explore various scenarios. This approach has been called
companion modeling (Bousquet, et al. 1999).
3 MODELS
3.1 Conceptualisation
The first model we designed was initiated during a meeting in the field in
September 2001 and discussions with the scientists involved in the project.
The research question identified at that time was: what are the socio-
economic impacts of different water control scenarios? The hypothesis is that
there are two main driving forces: the price of rice and shrimps and the water
quality (salinity).
IRRI scientists already produced a model to control the water dynamics
(Hoanh, et al. 2001). This model is used to assess various scenarios of
management of the sluices. It has been validated and is used on the field: for
given needs the model helps the management of different sluices. The main
potential input of multi-agent modelling is the representation of decision-
making process. The bio-physical model does not take into account the
dynamics of the stakeholders decisions and the heterogeneity of the
stakeholders. Thus we started to conceptualise a model to take into account
the decisions across different scales. Decisions are made by different
stakeholders at the level of the Province, the district, the commune, the
hamlet and the farm household. There are also other stakeholders who
influence the systems like the water company (linked with the Province), the
3
bank or different kind of traders. The purpose is to use the tool to understand
the coordination and explore different scenarios.
Among the various choices we made, we decide to design firstly a generic
model for Bac Lieu context. The purpose is not to represent exactly such or
such reality but to capture the interactions between the main processes from
which emerges the complexity. The assumption is that, once these processes
captured, the model can be conceptualised anywhere in the study area.
We made different choices:
• Biophysical representation. The spatial model is represented in raster
mode. Each cell represents 0.5 hectares: a farmer plot is thus
composed of one to 8 cells (0.5 to 4 hectares). In the model each
farmer owns only one plot. The model represents main canal, primary,
secondary and tertiary canals. The acidity of the soil has been
classified in three groups high land non-acid surface (high), medium
deep acid surface soil (medium), low shallow acid surface soil (low).
Sluices at the intersection between main canal and primary canal are
open or closed. Data are available to simulate the diffusion of salinity
depending on the state of the sluices. Each plot is connected to the
closest canal.
Secondary canal
Tertiary canal
Primary canal
Sluices
Main Canal
Figure 1. The interface of the first model,
Representing the plots, the canals and the sluices
• Farmer’s decision-making process. There has been several
discussions on the decision making process of the farmers. We
decided that the farmer agent in the model would represent the level of
the household. Each individual is not represented. Finally we worked
out the diagram presented in figure 2. The decision depends on three
parameters: did the rainfall come early or not? It is assumed that the
water company will open or close the sluices accordingly. What is the
4
topographic position of the plot? And finally what are the economic
conditions of the farmers? Depending on the combination of these
three parameters, the household will have different kind of activities on
his plot. It also depends on its access to credit. Along the year the
farmer will change crop according to the loans he could or could not
get.
Figure 2. The flow chart of the crop decision of the household agents
• Other agents’ decisions for water management. Two agents are involved :
the district and the water company. From the observations on the field the
commune and then the district compile a calendar. This calendar is
transmitted to the water company, which operates the sluices accordingly.
• Traders. The rice and the shrimps produced are sold to traders. As one of
the driving forces is the price, it is important to capture the set of
interactions for the two commodity chains. The model of interactions for the
rice is the following: the trader proposes price depending on request
(quality demanded), the trader has spatial zone and social relations (social
network), the trader has a limitation of volume (boat), the household
chooses best price. For the shrimp there are also middlemen who buy the
shrimps, depending on their quality.
5
3.2 Questions
The conceptualisation is an on-going process. Several questions are unsolved
yet. Some are related to the objective of the model. Some are related to the
knowledge we introduced in the model.
Some questions are related to the purpose and the use of the model. For
instance, the farmers’ decision-making process presented above is not linked
to the dynamics of the water salinity. It is assumed that, depending on the
rainfall; the sluices will be opened or closed. Thus, at this level of resolution,
there is no need of simulating the water dynamics. There is no spatial
differentiation between the plots. When the sluices are operated it takes only a
few hours for the salinity level to diffuse. Therefore at a weekly time step we
can consider the salinity situation as homogenous. This means that a model
to study the economic differentiation does not need to be linked to hydraulic
model, but only to the decision of the district and water company to open or
close the sluices. At this stage of the process this conclusion has to be
discussed.
The model can be used as a framework to guide the gathering of data in the
different surveys achieved by the DFID project. We give some examples of
needed knowledge to complete this conceptualisation process :
• We need to understand better how the information is transmitted
across levels of organization, particularly in the ascendant way: how
are the upper levels informed of the decisions of the households and
how is it taken into account?
• How do the economic conditions of a household change? In the
household decision making diagram, the crop decision is dependent on
the economic conditions of the household. What is the set of rule,
which changes these conditions according to the cropping pattern of
the household? How detailed must be the model? We can try to
simulate the budget of the household; by using data we have for costs,
yields and prices. It is maybe not necessary to have precise simulation
of the household budget. Simple transitions rules can also be used
such as:
o If the household economic conditions are average and he invest
in shrimps, it has 66% of chance to become rich, 17% to remain
average, and 17% to become poor.
• How do the households loan money?
• The market of the land has to be introduced. When agents become too
poor they may have to sell their land. Other agents may in turn
increase their area. This process is important to study the economic
differentiation
6
3.3 Computer model
We started the implementation of the above described model. It has been
implemented with the Cormas platform using Smalltalk object-oriented
language. The time step of this model is the week.
We can create artificially a topology. In the figure we created 4000 plots
belonging to 4000 farmers. At initialisation 33% of the farmers are poor, 33%
medium, and 33% are rich farmers. The rainfall is randomly drawn. What is
presented below is just an illustration: only the decisions making of the
farmers have been implemented. The transitions from an economic state to
another are non-documented. The behaviour and decisions of other agents is
not implemented yet.
Topology Economic conditions
Figure 3. Two snapshots of interface at initialisation representing a point of
view on the topology and the initial economic conditions
(uniform random distribution).
7
The following results are given as illustrations of the type of model output.
One can observe which crops the farmers have chosen.
Figure 4. Snapshot of the interface during the simulation
It is also possible to define and observe some aggregated indicators. For
instance it is possible to observe the economic differentiation for the full set of
farmers or for specific set of farmers. For instance in the following figure,
again given as an illustration, one can observe the evolution of the proportion
of the different economic classes.
Figure 5. Example of indicators:
Differentiation of economic conditions for the full set of agents
8
4 PERSPECTIVES
This modeling experiment is an on-going process, which will be assessed in
November 2002. This paper was written as a presentation of the process and
a mid-term report. The objective is to stimulate discussion. After this
preliminary work, the perspectives are:
• To complete and validate the conceptualisation of the model presented
above. The decisions making process of the households has to be
validated, and we need to input knowledge on economic dynamics and
social decisions for water control.
• To implement the conceptual model in Cormas software. This tool also
allows to import data form GIS. For instance we can use the map of soil
quality, as this is the main biophysical parameter in our model (figure
6). Thus, it will be possible to simulate the model on landscapes similar
to reality and validate the model by comparison between simulated
data and data from the field.
Figure 6. Soil Characteristics
• To define the use of such model. As any simulation model, it can be
used to explore different scenarios to enhance the understanding of the
scientists and help them communicate their results. It is possible to go
further. We try to use these models in what we call companion-
modelling approach. It means that we try to develop and use these
models with the different stakeholders to help better decisions and
coordination. For that objective, we have developed a methodology
based on the coordinated use of role games, interviews and
simulations with the stakeholders. This allows discussions among
stakeholders and the model can be used to help decision makers at
different levels. This mediation approach can be used on the Bac Lieu
case study. It requires the agreement and participation of the institution
in charge of the decisions, the appropriation of the methodology by
local and legitimised scientists and their medium –term involvement to
follow-up the process and its impact.
9
5 REFERENCES
Barreteau, O. and Bousquet, F. 2000. SHADOC: a Multi-Agent Model to tackle
viability of irrigated systems. Annals of Operations Research 94:139-162.
Bilderberg. 1999. Integrated Natural Resource Management, The Bilderberg
Consensus. in.
Bousquet, F., Bakam, I., Proton, H. and Le Page, C. 1998. Cormas: Common-Pool
Resources and Multi-Agent Systems. Lecture Notes in Artificial Intelligence
1416:826-837.
Bousquet, F., Barreteau, O., Le Page, C., Mullon, C. and Weber, J. 1999. An
environmental modelling approach. The use of multi-agents simulations. Pages
113-122 in F. Blasco and A. Weill, editors. Advances in Environmental and
Ecological Modelling. Elsevier, Paris.
Bousquet, F., LePage, C., Bakam, I. and Takforyan, A. 2001. Multi-agent simulations
of hunting wild meat in a village in eastern Cameroon. Ecological modelling
138:331-346.
Castella, J. C., Boissau, S., Lan Anh, H. and Husson, O. 2001. Enhancing
communities' adaptability to a rapidly changing environment in Vietnam uplands:
the SAMBA role-play. in J. Suminguit, Caidic, J, editor. Sustaining Upland
Development in Southeast Asia: Issues, Tools & Institutions for Local Natural
Resource Management. SANREM CRSP/ Southeast Asia, Makati, Metro Manila,
Philippines.
Ferber, J. 1999. Multi-Agent Systems : an Introduction to Distributed Artificial
Intelligence. Addison-Wesley, Reading, MA.
Gilbert, N. 1995. Emergence in social simulation. Pages 144-156 in R. Conte and N.
Gilbert, editors. Artificial societies. The computer simulation of social life. UCL
Press.
Hoanh, C. T., Tuong, T. P., Kam, S. P., Phong, N. D., Ngoc, N. V. and Lehmann, E.
2001. Using GIS-linked hydraulic model to manage conflicting demands on water
quality for schrimp and rice production in the Mekong river delta, vietnam. Pages
221-226 in F. Ghassemi, M. McAleer, L. Oxley and M. Scoccimaro, editors.
Modsim 2001. The Modelling and Simulation Society of Australia and New
Zealand (MSSANZ),, Canberra, Australia.
Rouchier, J., Bousquet, F., Requier-Desjardins, M. and Antona, M. 2001. A multi-
agent model for transhumance in North Cameroon. Journal of Economic
Dynamics and Control 25:527-559.
Weiss, G., editor. 1999. Multiagent Systems : a Modern Approach to Distributed
Artificial Intelligence. MIT Press.
10