Nils FERRAND
GIS, Spatial Planning, Multi-Agent Systems, Negotiation, Decision Support System, Multi-Criteria Spatial decision Aid, CSCW, Simulation of Societies
Large and medium scale spatial planning is very complex, regarding either its object : the project and its environment (ecosystems, landscapes, socio-economics, etc.), or its process, which implies many actors, with different world's representation and related interests, and who are individually attached to a specific territory. Issues like exchange and coevolution of spatial representations between many distant actors, negotiation support and simulation, multi-actor multi-criteria decision support for continuous spatial planning, are usually not addressed by current GIS. We propose to use Multi-Agents Systems (MAS) to enhance or develop such functionalities. We present two approaches: in the first we use Multi-REACTIVE-Agents Systems to solve the complex spatial optimization problems encountered in the search for least environmental impact area for infrastructures, where environmental sensitivities, structural constraints and different localized actors decision systems are used. In the second approach, we use Multi-COGNITIVE-Agents Systems to support and simulate the exchange and dynamics of spatial representations and policies, considering the general political values, the specific spatial constraints, and the socio-relational characteristics of embedded actors. We conclude on the general interest of using MAS for spatial modeling, and argue ab out the development of specific MAS based GIS.
As we are more interested in real world problems than in computer issues, we will address here a social and political process which is extremely important for the organization of the territory: spatial planning in its multi-actor dimension. We propose an overview of this process, which will show to be very dependent on the representation of space by the embedded actors. So our modelling deals with how world is perceived and communicated, and not how it is really. This perspective seems to be more operative in terms of decision making. But it is opposed to usual GIS principles.
In order to implement related information systems, we propose to use Multi-Agent Systems. We describe them quickly, and then present three applications: the first for spatial project analysis, the second for spatial negotiation support, and the last for spatial negotiation simulation.
Spatial planning is nowadays an extremely sensitive issue, especially in Europe, where old traditional settlements, high population densities within some critical areas, and complex political influences and relations, lead to a situation of permanent crisis regarding the use and destiny of lands. Environmental awareness since the mid 60's has brought new issues and related procedures, which tend to widen the potentially embedded system of people and constraints. Consequently, the process of spatial planning is hardly controlled, highly decentralized, and for wide area and major projects, can last for decades. Furthermore, the rationality of final solutions is not always obvious, regarding either social or environmental aspects.
So, being stated the importance and difficulty of these processes, it justifies the many studies and developments of dedicated information systems which were made. But among those, only few address the whole problem, including environmental assessment, spatial multi-criteria decision making, expert practices, multiple parties, consensus query, and negotiation. So, in this first part , we will try to analyze and describe precisely the processes to be supported. That way, we expect to be able to specify and develop dedicated information systems which could fulfill the different needs in an integrative way. Our reference here will be the french case, which can be somehow different from some countries, because of the extremely distributive (conflictuous, rather than cooperative) nature of the process.
Spatial Planning is essentially a decision process, which means that it includes at least the four classical (Simon, 1977) stages : intelligence (information query), design (of solutions), choice, and review (execution and feedback). But it has also other features :
So, three main systems are concerned: the environment (the territory and related elements), the project itself, and the set of embedded actors. Now, it appears that practically none of this real systems is operative; because as part of a decision process, only the representations of the environment, the project, and the society, by the effective decision makers (those participating in the applied decision), can induce reality at the end. This assumption relies widely on the proposals of the constructivist school of Palo-Alto (Watzlawick, 1978). It implies that spatial planning is totally dependent on its social and relational context. So it is almost impossible to modelize spatial planning without a model of social systems and its related decision making process. Furthermore it should be considered that spatial planning is in fact a co-evolution or a co-construction of shared and personal representations of space, which converge through the game of interpersonal relations. Thus, even if basically representations are built on top of mediated reality, it is fundamental to remark that in the perspective of modelling and supporting spatial planning, we will have to assess how spatial representations are built or exchanged and evolve. Such knowledge could also help to promote the cognitive ergonomy of our systems. Finally this conception of spatial planning as a social construction of reality (the territory), emphasizes the contingency of information, and bring the relativity of information as a constant background. To summarize : there is no data in spatial planning, only models.
In the following, we will distinguish three levels : the first deals with the space itself, and the state of the environment; it consists in analyzing the potential impact of the project, and proposing alternatives. The second level is linked to the social and political context; and address the nature, and the construction of representations of the space and the project. The third level is reached when representations are exchanged, and when negotiation occurs.
As an instance of the "intelligence" stage of Simon, the first part of the planning process consists in making what is usually called "Environmental Impact Assessment" (EIA, or EIE in french). Such work is realized by experts, chosen after their proficiency in the different domains to be considered (and the price they propose...) They are independent people, specialized in this activity. Their role is mostly to commit themselves into the conclusion they give. In fact the technical or scientific content of their conclusions is usually not at stake. And finally the decision makers, who are politicians, or administratives, just use experts as producers of "packaged solutions" to be compared. In the regular methodology we refer to, their specific work can be divided in three subphases (after the request by initial demander):
Next is a phase of integration of the different maps of sensitivity in order to propose a set of alternatives, with attached information about each of them. But, in a clean decision process, this last phase can't be done by the expert, as it includes a mixing of different criteria, which assumes in some way that a weighting is chosen. And such ranking of criteria is of the political domain. So this phase of "design" is no more directly related to the environment and the project; it is already a part of a social and political process.
As we stated before, spatial planning is a multi-actor process which cannot be separated from its social context, and which consists in the construction and exchange of representations of the space and project between the actors. So, the first question is to determine what representation (model or image) is used by each actor. Our assumption here is that the results of the expert analysis of the environment, and the proposed mono-criteria maps of sensitivity, won't be discussed as such. If so, then the topic is moved from the decision field to the scientific or technical controversy, which is not our subject here. Then we consider that the correct "model of representation" is a complex map, valid only for one actor, and giving in one hand the possible alternatives as thought by this actor, and in the other hand the perceived global (in terms of criteria) value attached to the different places in the area.
Actors can be extremely different, regarding to their interests, values, relational abilities, or spatial attachments. But in a first step, we will only consider differentiation of actors after their "political position", or value system. To model this we use the classical approach of multi-criteria decision theory (Vincke, 1989 ), which consists in weighting the different criteria. So, to have an idea of different representations or "points of view", we need to combine those weightings and the sensitivity maps as given by experts. We could get that way a spatial system of preferences attached to a certain type of actor.
To refine this, in a second step, we should include a tolerance system above the weightings of criteria. It would describe the acceptance by actors of a modification of their position regarding a topic.
Now, in a third step, the spatial situation of actors is to be taken into account. Obviously the political position of actors is not homogeneous in space. Some are officially responsible for a specific area, some have personal interest here or there (house...), etc. So, rather than a uniform system of weighting, we can use a mapping of weightings attached to each actor. In a way, such a map is a model of the spatialized system of decision of the actors.
The approach proposed here address only one single actor representation of space and project. But in the context of multi-actors spatial planning, the collective spatial representation of a group is more operative. So we need not only the collection of individual decision (political) systems, but also their combination. If we have a set of embedded actors, the collective representation will result from the collective decision system which reflects the exchange of power position between actors. This model is more difficult to build, and for a sake of simplification, we consider that each actor as a certain voting power for each area, and that by combination on each point of the votes coupled to individual weightings, we get the global collective decision map. This approach doesn't consider any type of negotiation at this level; it only deals with the fact that actors are spatialized, and that they are organized with a basic system of power. So the representation we get here is a rationalized image, eventually different from the real situation, which includes external constraints (brought by actors in a negotiation), and relational effects linked to the context itself of collective decision making. To conclude here, one can remark that we need an inventory of potential actors, and a mapping of their location. Such information is difficult to get, and some assumptions are usually made.
Given the incomes of the expert about the state of the environment, and some assessment of the decision system of embedded actors, with its collective counterpart, we can address the fundamental phase of negotiation. When actors meet to negotiate a planning project, they make proposals, exchange points of view, and argue based on spatial or overall issues. The evolution of the negotiation depends on what is said, how it is said, by who, to whom, in which context, etc. It's not the object here to dissert about negotiation, but it should be noticed that most of the arguments exchanged during planning negotiations are either directly of a spatial nature (location, extension, topology), or at least georeferenced. Furthermore the type of messages that are sent is such that most of their content can be mis-interpreted by the receiver, because of a lack of common knowledge or common representation of the space. Finally, following the assumptions made above, the objective of negotiation is to make actors exchange representations and let their own evolve until an agreement is found, which means that either a common spatial solution is found (consensus), or some actors are obliged to give back some of their pretentions, and will get or not some counterpart. Whatever is the perspective about negotiation, actors will have to exchange geoinformations.
After a brief description of the process we consider, we can propose a functional analysis of the related needs, and observe quickly how present technical approach fulfill the requirements. As a consequence, we propose some alternative technical answers.
As a principial for our whole work, we set that information processing systems should be developed according to established needs from the practice, and not as a proposed usage of already common systems or softwares. In our context, as we stated above that most of the information used in multi-actors spatial planning is models (representations), rather than data, then we should emphasize the dynamic and interactive construction and analysis of spatial models. The background idea can be expressed by: give the actor some bricks to build his world, and not a world to fit in.
What are the needs ? We give here a functional classification, which we will refer to later:
The different functional elements should be integrated in a coherent environment, dedicated to unskilled users.
Presently, each of the functions is addressed by the following domains or technologies (refer to the glossary for abbreviations):
Information | |
Modeling | |
Problem Solving | |
MultiCriteria Decision Aid | MCDA : ELECTRE, PREFCALC, PRIAM,etc. |
Negotiation | |
Communication / Publishing |
In our project, we will instrument only two functions: problem solving and negotiation. Because we estimate that the others are already filled by commercial products, whereas there is no answer for the complex integration of multiple spatialized criteria with decision systems distributed through space, and regarding negotiation, the principle of collective construction of spatial representation is not yet used.
We present thereafter the Multi-Agent approach and argue about its efficiency for our context.
The following is abstracted from a general presentation for spatial issues in (Ferrand, 1995 ).
The development of MAS followed two directions which fit the present main families in DAI . The first proposed to distribute problem solving between different modules, called agents, in order to distinguish different expertizes, to reuse common systems already developed, to mix heterogeneous processes, to gain openness and to get 'emerging functionalities' from the magma of interactions. It went through multi-expert systems with various organizations, and led to the present cognitive agents systems. This thread is mainly related to the symbolic approach, and it has shown its utility for solving complex problems. Cognitive agents have an explicit representation of their environment and of other agents. They have a memory; they build plans, selfish or cooperative; they exchange complex messages; and usually only few of them are used together in a system.
The second thread is more recent and can be called the reactive one. It has followed from many observations about information processing in nature, where 'intelligence emerge' from the interactions of multiple simple entities, acting on the base of direct reaction to stimuli. From this point of view, reactive agent systems are close to the connexionist paradigm, and many of its issues are shared. But the reactive paradigm also rely widely on ethology, where it is observed that extremely structured objects or processes are built by colonies ('swarms') of very simple animals, without intelligence, like ants or bees. In that perspective it is close to Artificial Life.
In our conceptual definition, a multi-agents system is a set of agents interacting in a common environment, where an agent is an entity living in this environment and able to modify both its environment (communication, decision, action) and itself (perception, reasoning, learning). The environment is composed of all the entities from the addressed universe which are not agents. So an agent can perceive and represent its environment, it can communicate with others, and it has an autonomous behaviour depending on its observations, knowledge and interactions. 'Autonomous' means mainly here that its behaviour is attached to the agent (self-control) so that whatever the social or environmental context is, it will have the same functional response to stimuli. Therefore no external control exists and impose an activity referring to informations and behavioral laws unknown by the agent. It should be noticed that such an agent can recursively be itself a system of sub-agents.
In the cognitive case, there is an explicit and symbolic representation of the environment and other agents. Cognitive agents handle goals, plans and resources. They have an expertise expressed in symbolic terms. Their communication with others is managed explicitly and the exchange of messages follows some fixed protocols. The reference domain is more or less sociology.
In the reactive case, there is no explicit representation of the environment or other agents, no memory, and the behaviour of agents is based on sensory-motor functions. Such agents are in fact complex automata with the ability to control their interaction pattern. One big interest of the reactive approach is that with such a simple behaviour, one can expect that some formal results about the global dynamic of the system could be obtained.
The implementation relies on the design and building of code modules that fit each agent. Present usual approach consists in using "Object Oriented Languages" and dedicated environment or class libraries, but this choice is up to the designer. The main issue in the implementation is the management of parallelism. It can be real parallelism, if possible, but more often it is simulated parallelism, either with UNIX processes (using PVM for instance), or simply with regular synchronous or asynchronous iterations.
There are two possible use of Multi-Agent Systems. The first one is simulation, where a computer model of the real world is built and used as a laboratory to study some process. The second approach is problem solving. It deals with the same issues than spatial optimization, which is a very classical problem in AI. To summarize, it tries to find a shape or pattern in space and time which fits some constraints.
A Multi-Agent based Simulation consists in mapping the agents onto the modeled objects of the observed system. The closure, the interactions, and the initial state are also identical. Agents' state is defined by characteristic attributes, including at least its position, but also all other changing parameters. In the classical example of ecological simulation, agents would be all the modelled beings, with their position, species, age, sex, size, health, etc, as attributes, and their feeding, reproduction, moving mechanisms as processes. The environment would be the hypsometric model, the vegetables, the weather, seasons, etc... Many agents can be created and let evolve in their environment.
As stated above, the issues are the same than those of classical optimization. But there, an utility or 'energy' function is defined for the pattern, which comply with the constraints, and a solution is found using minimizing methods . Whereas in the MAS paradigm, the pattern is divided into agents which will interact with the data (their environment) and with the other agents. As soon as the problem is concerned with two dimensional patterns (lines, zoning, networks), this approach is very relevant. The division of the pattern in agents leads to a set of subproblems localized in space, which can be autonomously solved. After their construction, agents evolve autonomously, and in parallel. If it doesn't stabilize, it means that no solution was found. Then the constraints have to be relaxed, or some parameters, like the agent sampling of the pattern, must be changed. If it stabilizes, than the reached pattern is a solution as it means that each agent (hence each part of the pattern) fulfills its requirements both versus the data and versus the pattern. If another solution is required, the system has to be reinitialized in a distant state. In other words, any steady state is the expression of a solution to the problem. This approach allows several variations: the problem can evolve with time; it can depend on the position; different problems can be combined easily; the stability of the solution can be analyzed just by changing some parameters and letting the system reevolve; qualitative and uncertain constraints can be used.
The use of MAS supposes to come back to a naturalist approach, because a description of the objects and processes is needed. But a qualitative analysis is possible as the computer model support such processing. In particular, symbolic rules can be used. According to the experiences we have made in relation to expert modelers (Ferrand and Demazeau, 1994 ), the type of modelling we require is very simple and natural to skilled experts.
We can accept sparse information as soon as they are connected in terms of processes. Furthermore, any type of information can be integrated : rules, functions, differential equations... And finally the model is entirely open: it is possible to add agents of any type at a low cost.
The modelling procedure is a classical systemic approach:
Using a Multi-Agent implementation, it becomes operative.
We present thereafter three different systems, based on the Multi-Agents approach, which are applied to Multi-Actor Spatial Planning. The first is dedicated to spatial project analysis , where multiple criteria are used, and multiple actors with different points of view are embedded. The second is a project which deals with the support to negotiation of spatial projects and exchange of geographical informations and other type of arguments between different actors meeting on electronic networks. We introduce a specific agent based approach by implementing a mediator agent concept. The last is a tentative simulation of spatial negotiation , considering environmental constraints, political choices of actors, relational effects, and contextual dynamics.
After a long time spent in observing the practice of environmental assessment experts, and with a common analysis, different problems appeared which were mainly related to the phase of design , consisting in integrating environmental sensitivity maps into a proposal of alternative solutions:
The implicit request was to build a system which could integrate the sensitivity maps given by the experts, the structural constraints of the project, and a political position (an actor description) expressed by a weighting of the criteria, into a map of least impact solution for the project. Further extensions include spatialization of decision systems, sensitivity analysis of different solutions, and proposals for consensual alternatives.
Our answer was the SMAALA system (Systeme Multi-Agents d'Aide a la Localisation d'Amenagements, or Multi-Agent System for Support to Infrastructure Localization), initially developed on PC, and now transferred to UNIX SUN stations and Macintosh.
The application field retained for the first version of SMAALA is the planning of linear infrastructures (roads, railways, electric lines) through territories. Formally, such planning is mainly an optimization like problem, but multi-dimensional and with spatialized constraints and decision system (similar to spatialized optimization process). Because of a structural similarity of the problem with some former experience of complex image analysis, and robotics path planning, we decided to use Multi-Reactive-Agents Systems (MRAS) to solve it. The choice was motivated both by the accuracy of MRAS for complex spatial problem solving, because of the structural analogy of the problem and its processing (Ferrand, 1995 ), and by the fitness of the MRAS model to the mental declarative models used by the experts.
SMAALA was developed in 6 phases:
The first level of the system is the pure least impact path finding, for one decision system, uniformous on space. SMAALA accepts the different sensitivity maps given by the experts, one for each criteria. It takes also as an input the morphological constraints of the linear project (curve, slope), and its topological constraints: start and arrival points, off-limit area, etc. Finally it uses a weighting of the criteria, simply given as an n-dimensional simplexes, and which describe then decision system of one actor. According to some process parameters which set the precision, SMAALA gives the different solutions for this actor, as a set of path, following the constraints, and minimizing the overall impact according to the relative importance of criteria given by the weighting. Any type of further spatial analysis (length of the path, statistics of impact) can be processed with classical GIS tools.
The principle of the optimization can be explained quite simply by the following metaphor: let assume that we plan an electric line between two points, and that each pylon is active and can move. It has perception ability toward its environment, and it is linked with other pylons by the line. They constitute the set of reactive agents. The problem is to find a position for each pylon so that the global impact of the line is minimized and the structural constraints verified. Then the pylons / agents individually will solve a very simple local problem which is: 'where should I go in my close neighborhood to decrease my impact ?' And then: 'Ok, but can I go there considering what my neighbors do ?'. So starting from a random choice (or straight solution) we let the system evolve (synchronously) till it gets satisfaction. To compute the local decision, we use the weighting as a parameter, and accept a direct normalized aggregation (utility based local decision making). By iterating different initial position, we finally get the collection of attractors of the system which show the solutions. Solutions can be given with a width showing the relative size of attractors (area of equal impact).
One can remark that regular Multi-Criteria methods (utility based, or overranking) cannot be used because of the continuity of the solution space. However, at this step, usual optimization (optimal path finding) is also possible, as the decision system is homogeneous and isotropic. But with the MRAS approach we allow the use of parallel processes directly, and facilitate the analysis of the proposed solution through dynamic modification of the parameters (including weightings) with fast reconfiguration of the solution.
To summarize, this level gives the representation of the solution to a planning project for one actor who considers all the space the same way. It should be noticed that the system usually gives only a set of possible path, and not the best one. The final choice remains to the decision maker (eventually with usual tools like ELECTRE).
The second level includes the fact that decision systems of actors are spatialized: for instance the major of a town as not the same point of view (political position) in its town, and outside, and if he has other interests in some area, he can have still another position there. Then, instead of using a single set of weightings for the criteria, we replace it with a map of the position of an actor. For each point in space, there is a different weighting. Furthermore, it is established by experts that the least impact path finding process is anisotropic: the impact of an electric line following the main direction of a valley is reduced regarding to the same crossing the valley. Such parameter is implicitly integrated by human experts during his assessment. So the local aggregation result also depends on the local direction of the linear solution.
To implement all those constraints, we ask that the agents (pylons in our metaphor) use a spatialized and contingent local decision system: a different type of optimization for each point in space, and each direction of the current state of the solution. We suspect that such constraint is difficult to handle through usual analytical optimization process.
The third level address the sensitivity of the found solution to a change in the actor's position: what happens in term of representation of the project by the actor if he changes his mind regarding general balance of criteria, or otherwise, what is the acceptability of modification of his proposal ? This question is very important because it prepares the ultimate negotiation by assessing the "open doors".
It is quite simple to get this result. In a first step we get the solution with a first system, and record it. At this stage, agents are localized in space as a sample solution. If we perturbate the weightings, agents will adapt automatically their position to fit the new constraints, and then give a new solution to compare with the first. The only difficulty is to choose a perturbation process in the n-dimensional criteria space.
Having an image of the preferred solutions for each actor, we can compute the least conflictuous (least risky) solutions. Such solutions are the least sensitive in terms of decision system. Their computation is expensive, because the process consists in sampling the possible decision system with a common step (building a grid of values of criteria) and applying them to the SMAALA regular finding. Then for each type of initial weighting, corresponding to every possible type of actor, a set of solutions is found. By iterating this process, and selecting each time the least sensitive solutions (with a threshold and a limited number of paths), we get a set of expected least conflictuous solutions. It is useful when embedded actors are unknown before the negotiation.
In practice the number of observed solutions is limited, whatever is the context.
A rational compromise solution can be sought when the actors are known previously. All the solutions (representations by actors) are computed, and by superimposing maps we directly get the common paths, if there are. At least some parts are common, and can be used as a start for discussion.
This last level adds a relational dimension, which is more subject to discussion. The idea is to take into account the fact that actors have different power and voting position. We have different possibilities: put a weighting on the global decision system of actors, give a priority to some actors for specific criteria (notion of expert), or even spatialize the relative importance of actors, each one having its own "territory". All those approaches are possible, and suppose to require that agents (pylons) use a pre-computed local system of decision to integrate the different environmental values. We obtain the collective solution that is potentially found, without any consideration of relational effects, or contextual negotiation processes. It assumes that enhanced actors knowledge is available.
This second application belongs more or less to the Computer Supported Cooperative Work (CSCW) thread. Its objective is to support negotiation of spatial planning between distant actors using electronic networks to communicate. It is a current project (Koning & al., 1995 ) developed within MAGMA group as an instance of its specific Multi-Agent approach, and implemented mostly in the JAVA language from SUN . There are two main aspects: the first is the possibility of collective work with shared geographic informations, exchanged through networks like Internet. We won't address it here, as there are already strong related programs (OpenGIS , GEOMED ). The second aspect is the possibility to facilitate and "mediate" negotiation about spatial issues. In this prospect, a proposal was made, based on Cognitive Agent Systems. The main principle is to attach to each actor an agent "assistant" or "mediator", who knows its actor and the other assistants, and tries to help the process. It is not the object here to address the pure CSCW aspects. So we focus only on the geographical aspects of the mediator. Its main related duty is to handle georeferenced arguments in order to control and illustrate the spatial organization of issues. For instance mediator should provide the actor with any type of information which would be spatially related to a presently evaluated proposal. If a road is at stake somewhere, the mediator should give its actor info about any close object, or look for regulations about neighborhood of roads, or enquire about possibly interested actors. Meanwhile then assistant should also check the spatial coherency of arguments. For instance if a road is not to be less than 100 meters to a river, it should check that no argument break this rule.
To precise the function and content of the assistant we distinguish the following parts in the cognitive agent architecture:
This architecture is quite a classical Multi-Agent one, and is implemented using a regular platform containing packages for Agents, Environments, Interactions and Organizations. But in this first stage, we don't stress the presentation and interface aspects. So, maps are still not used as a medium, and we are expecting JAVA applications handling the exchange and manipulation of maps.
As we are interested in supporting multi-actor spatial planning, we would like to be able to prepare negotiations and foreplan different strategies. In that perspective, and as there is no strictly rational approach, we try to develop a simulator of spatial negotiation, which could be used as a virtual laboratory. Negotiation as a whole is already much studied (Raiffa, 1982 ; Dupont, 1990 ), but spatial project negotiation is not usually addressed, as it much more difficult to discuss about a continuous space of solutions, than among a finite set of alternatives (as in usual business negotiations). Furthermore, in the context of multi-actor spatial planning, the closure of the system is extremely far. Many actors can demonstrate their interest and appear suddenly in the process. So it is mainly an open negotiation, very difficult to limitate. And the behaviors of actors are very complex, with different roles and hierarchical levels.
The analysis leads to two levels: the first is conceptual and address the modeling of actors as individuals, the second is instrumental and address the implementation of such a model into an effective simulator.
The model of planning actor we propose includes:
Two dynamic perspectives are available: in the first a spatial negotiation is a sequence of proposals (paths in the case of linear infrastructures) which are evaluated by the different actors, and some counter-proposals are made, until a final consensus or compromise is reached. It is the contingent approach. The second is the structural social approach (or strategic game approach), where actors consider the object of the negotiation only as a justification and a base for settling prestructured interaction systems, where they try to maximize their global gain in the process. This last is probably the most accurate to reality, but it is much more difficult to implement.
The SPAtial Negotiation Simulator (SPANS) we develop after this model is based on a multi-cognitive-agent approach, in which each modeled actor is implemented in an agent (straight fit). We use all the domain of DAI background knowledge to facilitate the dynamic processing of agents interaction. The approach retained in a first phase is contingent. The core idea is to use for each modeled actor a map showing his representation of the solution. This map is not shared. There is also a shared map which shows the different proposals and their overlapping. So when a new proposal is made by some actor, each of the others evaluate it first according to his value system, which is absolute. It gives a political assessment of the proposal ("this proposal damages the environment in this area, but it is acceptable"). But it is balanced by objective constraints ("Oh, it cannot pass through my backyard"). And finally relational elements are considered ("Well, he is my boss..."). At the end of this evaluation, each tough point is marked and a commitment is put on the counter-argument. The foundation of the proposal system is to "handle open doors", which means that permanently the objective of actors is to keep themselves open choices within space, and to prevent others from being blocked. This is why we expect to reuse the SMAALA system as an evaluator of the possibility of changing a solution, according to each actor. Because "having an open door" implies that in the present situation some changes could be accepted without disturbing values or constraints.
SPANS is still in development and it should use the same geographic interface than the mediation support system.
After an analysis of multi-actor spatial planning, which insisted on the contingency of the exchange of spatial representation at stake, we proposed here an inventory of related functional needs. According to it, we shown that at least two ways could be followed. The first concerns spatial problem solving, and is dedicated to least impact area finding for spatial project, in the context of multi-actor planning. The second is related to negotiation, either for supporting it, or for simulating it.
Because of the structural complexity of these issues, and for a sake of cognitive ergonomy, we proposed to use as an integrative paradigm the Multi-Agent Systems. In that perspective, we shortly recall their principles. And describe the three projects we have in this context. The first, more advanced, deals with complex spatial optimization, with multiple criteria, and spatialized decision system. The second shows an instance of agent based CSCW system dedicated to spatial negotiation support. The last, more prospective, build a coherent framework for simulating spatial negotiations. In all those projects two background principles are kept: consider the social and relational issues as fundamental in planning processes, and build the information processing as an answer to recognized needs. Consequently we refuse the reduction of the complexity of problems. But we also show that Multi-Agent Systems can handle most of such problems, as they are themselves intrinsically complex.
So, as a conclusion, we would like to argue in favor of the development of a fully "Multi-Agent oriented GIS", in which any spatial entity would be implemented in an agent, which, according to the context, could be either active or passive. Such approach would have different advantages:
According to the present trend toward active object oriented GIS, we presume that our wish will soon be reached. .
SMAALA and SPANS project are supported by a grant (1993-1996) by the Region Rhone-Alpes , with the partnership of CERREP S.A., a consultancy agency specialized in land-use planning and environmental assessment.
Best thanks to Dominique MICHELLAND, Yves DEMAZEAU, Pierre DUMOLLARD and all the MAGMA / LEIBNIZ team.
This document was built using Tamaya-Thot from OPERA project , © INRIA -IMAG 1995.