Randy H. Gimblett
Department of Landscape Architecture
College of Architecture and Planning
Ball State University
Muncie, Indiana
George L. Ball
Advanced Resource Technology Program
University of Arizona
Tucson, Arizona
.
The use and popularity of Geographic Information Systems (GIS) has radically increased over the last decade and ultimately changed our perspective of environmental planning and the world around us. Traditional modeling approaches however provide only a static representation or effectively a time slice view of the situation. In addition, a large percentage of the knowledge inherent in a real physical/biological system is dependent upon the spatial association of the system components. To predict dynamic changes over time within an ecosystem, models need to maintain the spatial relationships found within ecosystems, and they must adequately model and incorporate animal or human behavior. This paper describes a new approach to modeling animal and human behavior in the context of actual environments referred to as an "Intelligent Action Model". The Intelligent Action Model for this type of simulation is controlled by a hierarchical neural network which provides both primal motivations and adaptive learning. The intelligent action models are created by linking animal behavior into existing dynamic physical models through a genetic algorithm classifier system and a simulation tool referred to as PROMAP. Models developed under this framework will provide new research tools for modeling a wide range of environmental problems such as fire, hydrology, pest management, wildlife behavior, and human perception. These techniques will undoubtedly lead us well into the 21st century for modeling and simulating spatial dynamic processes of natural systems.
While a tremendous amount of research over the years has focused on modeling ecosystem processes, very little has successfully achieved this task. This is due in part to the immensity of the data bases required for adequate description of the system under investigation and sophisticated data management systems for searching, assessing and manipulating such data. In addition, the transition functions used in the model must be able to handle spatially oriented data and provide a number of spatially oriented solutions.
This paper describes a framework being developed for building spatial dynamic models of natural processes in which an essentially unlimited number of environmental variables can be used to describe actual geographic areas. Problems associated with large data sets, as well as the need for hierarchical interactions have been addressed through the incorporation of various aspects of artificial intelligence. The work described here is an ongoing project and certain aspects such as the artificial intelligence components are still being developed. Deterministic spatial dynamic models are already being implemented and this paper will provide a background on how these models are designed. In addition, we lay out the work in progress as an indication of the potential of this type of modeling for researchers in other fields.
In the following sections we will outline how the environment is described; how single processes and multiple processes can be handled; how hierarchical interactions can be incorporated; and how it will be possible to include the influence of animal behavior into the models.
An ecosystem is a collection of interacting elements that together make up what is called the landscape mosaic. Not only are the component elements important, but also their arrangement within the mosaic. Data bases which describe this mosaic can be found in what are called geographic information systems or GIS.
Burroughs (1986) describes geographic information systems as"... computer based tools designed to collect, store, retrieve and change at will, manipulate, and display spatial information from the real world for a particular set of purposes."A GIS describes objects from the real world in terms of: Their position with respect to a known coordinate system; Their attributes, such as elevation and soil type, that are unrelated to position; Their spatial interrelationships with each other (topological relations), which describe how they are linked together or how it is possible to travel between them.
A GIS consists of several layers or maps, each of which describes the spatial distribution of some attribute, such as vegetation.The data bases can describe an area less than a square meter, or greater than several hundred square kilometers depending on the intended analysis. By the use of appropriate operators within the GIS program, questions related to the area described by the data base can be answered.These questions usually take the form of a suitability analysis which is the intersection of a given set of criteria within the data base. Each intersection indicates geographical location which meets the defined criteria.
The amount of information contained in a GIS is limited only by system storage capacity and the availability of attribute data.Storage requirements increase with increasing resolution of the data. An adequate GIS data base would allow the researcher to investigate for example, how changes in vegetation cover would effect the sedimentation rate of a local stream. This type of inquiry would require that information be passed across the data matrix in a manner consistent with known processes.Describing processes as spatial transitions imposes restrictions on the data storage method used in the GIS.
In vector systems the information about attributes is stored as polygons described by connecting nodes. In raster systems the information in each cell of a matrix describes the attribute for that geographical location. Processes are described as effects within neighborhoods and vector based GIS is not suited to neighborhood analysis. For that reason, raster data bases are used in this framework.
While GIS is an excellent means of depicting the environment, much of the applied research to this point have been simple suitability analysis, using a limited set of variables and map themes. While this approach to modeling environmental relationships certainly has some utility for a variety of applications for both local and regional government, it none the less is static representation of the environment or in effect, one slice in time.
For a model to simulate a natural ecosystem it must first be able to represent spatial attributes found in actual landscape (structure). Second, the model must be able to handle the flow of objects within the landscape (function). Third the model must be dynamic rather than static (change). Fourth, the model must allow evolution through adaptation. Each criterion links together to form the hierarchical structure that is necessary to complete the model.
The problem from a modeling standpoint becomes a matter of cellular interactions describing both the transitions between cells for a single process and hierarchical interactions.
Take for example a water runoff problem. Sediment transport between cells representing a watershed will result in transformation of the topography of the watershed over time. This is the result of a local action causing a global change. Describing water runoff over portion of the terrain within a cellular neighborhood will require identifying which areas are downhill, what are the soil characteristics and, what kind of vegetation cover is in each of the cells. Although this is an over simplified list it does point out the potential complexity of the appropriate algorithms. One factor that does provide a bit of simplification is the recognition that most processes can be described as cellular transition rules within the nine cell neighborhood. The process can be extended to larger areas by using the neighborhood as a moving window.
Two other similar problems within this watershed could be the reforestation of a watershed which would reduce the amount of runoff and therefore the total amount of sediment at the exit point of the watershed. The result of a local action causing a global change. Fire burns off the reforested area followed by heavy rainfall results in increased erosion and increased sedimentation at the exit of the watershed. The result of multiple process interaction causing both global (erosion effects hillslope profiles) and local (formation of channels and sediment buildup) changes.
While these may appear to be fairly simple problems on the surface, they are enormous complexity when viewed from a dynamic modeling perspective.
Attempts to use existing GIS programs to model spatial dynamic processes (Wilke and Finn 1988) have shown that the approach has promise but also problems. GIS programs were never intended to be used for this type of modeling. The algorithms used in the GIS operators were not designed for iterative processes, and in most cases are integer based. For this reason, researchers have devised specialized programs to work in cellular data bases. Most attempts have generally simplified the complexity of the environment (e.g. a lack of topographic data or other attributes) or the complexity of the algorithms (e.g. no diagonal transitions between cells for the sake of programming simplicity).
An alternative solution has been proposed by Sarennma et al. (1988) using an object oriented approach. The model developed is intended to describe the behavior of animals in natural habitats. The model obtained from aerial photographs, is classified into objects. The classification approach reduces the spatial information available to the system by combining areas of similarity into single homogeneous regions. Although all cell based systems assume uniformity within any particular cell, classification results in non-uniform sizes of regions. In addition, once within an object (e.g. a forest area), exact location can no longer be determined, and pathways between objects are predetermined and are not the result of actual terrain or other natural factors. This approach suffers from problems similar to those mentioned earlier in attempts to use GIS operators.
To overcome the limitations imposed by existing GIS programs and the arbitrary constraints imposed by other researchers, Ball (1990) designed and authored PROMAP. PROMAP is a research tool that allows the construction of spatial dynamic models using the data bases of raster GIS. It is a real number system with algorithms designed for iterative processing.
As an example of the capabilities of PROMAP, Ball (1990) implemented a model of surface fire spread originally designed by Vasconcelos (1988).The new version of the FIREMAP model was able to approximate the shape of a surface fire under specified conditions without initializing templates. Although work continues on the improvement of the FIREMAP model, it demonstrates that complex spatial models can be designed using this approach.
By adopting PROMAP as the means of manipulating the information in the data base, spatial dynamic models of complex processes can be developed. As with other spatial models, two problems still exist even in the models developed under this system: the lack of multiple pathways of action and; the lack of true hierarchical interaction.
The lack of multiple pathways is not merely an IF/ELSE situation.In natural processes, under certain conditions, two or more choices of action might each result in an acceptable solution. The ideal model would, over several starts, give approximately the same answer but very rarely the exact same answer. To achieve this type of model response, cellular transitions should derive actions based on current environmental parameters. A possible solution to the multiple pathway problem can be found in hierarchy theory.
A mechanism to allow hierarchical interaction within the model may be found in the work of John Holland. Holland, et.al. (1989) address the problem of levels of approximation in describing systems. They state that "...given the complexity of realistic environments and the limitations of realistic [modeling] systems, it is unreasonable to expect...models to be isomorphisms in which each unique state of the world maps onto a unique state in the model" (Holland, et.al.,1989, page 31). In their book, they consider the design of a layered set of transition functions which they call a quasi-homomorphism or simply a q-morphism.
The q-morphism is the mapping of many attributes in the real system through a hierarchical set of transition functions in the model. The adoption of Holland's q-morphic approach would satisfy both the requirements of hierarchical structure and multiple pathways. Linking the q-morphic approach with PROMAP could be accomplished through the use of two types of artificial intelligence techniques: neural networks and genetic algorithms.
Various artificial intelligence techniques have been developed over the years that attempt to solve aspects of learning dealing with pattern recognition, decision analysis, and optimization. Two areas that are relevant to the framework being discussed in this paper deal with data filtering and optimization of coordinated activity. Data filtering has been examined using artificial neural networks. For a basic understanding of neural nets the see (Wasserman, 1989).
To solve the problems of a lack of hierarchical structure, multiple pathways and adaptation, a new PROMAP operator has been conceived that incorporates intelligence as part of the operator's basic algorithm.This intelligent operator has been designedwhich is composed of two types of artificial intelligence techniques: neural networks and genetic algorithms.
The learning ability inherent in neural network technology enable the intelligent operator to examine and obtain data from multiple "sensors" and supply a set of weighted values that describe the current environmental conditions and biophysical content in a more concise manner. The action of neural networks is more than a simple recoding scheme. Nets must be constructed and trained for specific tasks and therefore the framework consists of multiple networks. The need for such complexity of data filtering is easily seen when the amount of data contained in a GIS structure is considered.
In much the same way that humans ignore (or suppress) much of the information around them on a routine basis, that is non-essential, good models which have access to huge amounts of data need to disregard information which is not relevant to the current problem. In addition, the information which is relevant should dictate a flow-of-control strategy derived, perhaps, from an ancillary knowledge base. One method of handling both these factors is through neural network architectures.
Within the structure of the model there are two levels of environmental information.First is the level of the global system. At this level the model is concerned with general conditions such as mean temperature. Variables that change gradually may appear static and generally effect the model at very large scales. On the other hand, the second or local level may change quickly, be influenced by global conditions, or be effected by neighboring events.Actions that are occurring on the local level need to have information which is relevant at the neighborhood scale. Neural networks are included to provide this type of information in a manner consistent with the requirements of the model. By accepting information supplied by specialized operators within PROMAP, the neural net would then return the filtered data to other operators within PROMAP.
Using neural nets as the primary control structure allows simultaneous multiple processes. The neural network would schedule certain types of events ahead of others based on a precedence level system. Simply, global actions need to occur prior to local changes (it should rain before a water runoff simulation is activated). Once the environmental information has been processed it must illicit some type of action. Action in the context of the model may simply be a change in temperature or perhaps the loss of trees due to fire. This implies the possibility of multiple actions and in a dynamic environment this may mean action based on data from multiple, possibly conflicting, inputs. Dealing with this situation in the modeling framework is the domain of the second use of artificial intelligence techniques: classifier systems which incorporate adaptive learning. These systems derive their power from the use of genetic algorithms.
Genetic algorithms (GA) are algorithms based on the mechanics of natural selection and natural genetics. GA have been shown to be effective in the areas of machine learning, search, and optimization strategies (Goldberg, 1989). In the complex spatial models being designed under this framework, all three areas are important.
GA provide a robust methodology for restricting the search for a "best" choice of action to those possibilities with the greatest potential for success. Success in this context should be thought of as a solution which results in an action consistent with some expected outcome. This does not imply an optimal solution, only that the simulated process behaves in a manner that would be observed in nature.
GA uses a population of "strings" that are used to examine current environmental conditions and arrive at some solution (action). Strings can be envisioned as chromosomes, and strings that provide better solutions are retained in the population at a rate according to a payoff scheme (fitness). To keep highly fit strings from overwhelming the population, natural genetic functions of crossover and mutation maintain variation within the population. The string population acts as the memory for the specific action that is being evaluated. If the component is capable of learning from its past attempts, the total system gains new knowledge.
A process may (and usually does) consist of multiple actions. A model could code an entire set of actions associated with a process into a single function, which is generally the accepted strategy of most modeling schemes. The drawback is the need to determine, apriori, what actions are to be done and under what conditions. This is the approach used with rule based systems and the weak point is known as the uniqueness factor.
For rule based systems to function, an assumption is made that all knowledge of the system is complete. If the model encounters a unique situation it fails. Real-time action in a dynamic environment virtually guarantees that situations will be encountered that have not been anticipated by the designer. It is this type of situation that GA provides an effective solution. Individual components are capable of adapting, allows them to respond to unique situations by making a "best guess". By dividing the process into smaller components, it only becomes necessary to activate the required components to achieve the desired actions.
Through the use of a "tag" system, the component operators can be triggered as needed through what is termed spreading activation. The component operators that have been activated invoke the needed hierarchical response within the model and the system.
The framework as described in this paper offers the researcher interested in ecological processes several advantages. First is the capability of using data bases which reflect real-world environments and have the potential for being tested. Building a model of an ecosystem process and running it in the laboratory is of no benefit unless the model can be tested against the real system. Some natural processes take too long in nature to be useful, but others are easily available to testing. The FIREMAP model is ready for field testing and can be applied to prescribed burns or to examine historical fires. Application of this framework for examining hillslope evolution due to water runoff as well as defining transitions between sheet flow and rill formation are under development.Other areas of potential investigation are population dynamics involving fluctuations in genes.
Second, this framework also provides the possibility for simulating animal behavior in complex natural environments. Along these lines, Wilson (1986) developed an artificial animal model (ANIMAT) acting in a simulated environment containing food and other objects. The environment was represented by a two dimensional array of hexagonal cells which were labeled as trees, food or, vacant. The ANIMAT's task was to explore its 18 by 58 cell environment consisting of woods and open space, finding food and avoiding trees. According to Wilson (1986) the ANIMAT's basic problem was "the generation of rules which associated sensory signals with appropriate actions so as to achieve the optimization [of the rate of occurrence of environmental signals]." Wilson's intelligent performance and adaptation algorithms are examples of the use of genetic algorithms. The cell based environment that Wilson uses translates directly to the structure used in the framework described in this paper.
Third, the framework also provides the ability to work with entirely synthetic environments or artificial worlds of any predetermined complexity. This aids the investigator by allowing algorithms to be tested under conditions for which the outcome is predictable. Being able to validate the function of individual components makes it much easier to validate the overall model.
Work has continued using the PROMAP system to develop dynamic models. Currently, four models are under development. They are: Fire Behavior, Hillslope Evolution, Forest Succession and Hiker Behavior models. While these models provide an opportunity to test the viability of dynamic spatial models, as of yet the intelligent action component has not been fully incorporated.
Two projects are currently underway to examine the viability of neural nets and genetic algorithms before they are structured into the intelligent operator, described above. A pilot study is underway to develop a neural network prototype decision support system for forest management applications. In addition, genetic algorithms are being incorporated into an artificial animal project in conjunction with the Ecology and Evolutionary Biology departmentat the University of Arizona for research and exploration of artificial biological specimens. Both of these projects will provide a basis for the development of the intelligent operator as described earlier in this paper.
The advantages of being able to use data bases that describe large scale, actual environments opens the door for producing models that can be applied to real problems. The spatial dynamic modeling that is described in this paper is not meant to replace the traditional mathematical models. What is presented here is an alternative that can be used to design and test ecosystem models in real world environments.
The model of surface fire can be tested under actual conditions to determine how well the model can predict the spread of the fire. In addition, the model can be applied to any area that has been mapped into a GIS data base.As the framework progresses, the inclusion of the artificial intelligence components will extend the modeling into multiple processes. Although some processes such as forest succession would not be testable, the ability to examine the potential effects of fire, rain, grazing and other perturbations in various intensities, would provide more insight into process interaction.
The ability to access existing data bases allows the modeler to enter the investigation of complex systems through the use of tools such as the framework described in this paper. Although this is a different approach than traditional modeling, it provides not only a means of examining ecosystem processes, but also of extending the usefulness of the models from research to application. This approach will undoubtedly lead us well into the 21stcentury for modeling and simulating spatial dynamic processes of natural systems.
Ball, George L. (1990). PROMAP: A Simulation Language for Spatial Dynamic Modeling. Ecological Modeling
Burrough, P.A. (1986).Principles of Geographic Information Systems for Land Resource Assessment. Clarendon Press, London
Goldberg, D.E. (1989). Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Publishing Company, Inc. New York, New York.
Holland, J.H., K.J., Holyoak, R.E Nisbett &P. R. Thagard. (1989). Induction: Processes of Inference, Learning and Discovery MIT Press. 389 pp.
Saarenmaa, H., N.D Stone, L.J Folse, J.M Packard, W.E. Grant, M.E. Makela & R.N.Coulson, . (1988). An Artificial Intelligence Modelling Approach to Simulating Animal/Habitat Interactions. Ecological Modelling, 44: 125-141
Vasconcelos, Maria J. (1988). Simulation of Fire Behavior with a Geographic Information System. Unpublished Masters Thesis, University of Arizona.
Wilke, David S. & J. T. Finn. (1988). A Spatial Model of Land Use and Forest Regeneration in the Ituri Forest of Northeastern Zaire. Ecological Modeling, 41:307-323.
Wilson, S.W. (1986) Knowledge Growth in an Artificial Animal. In: Adaptive and Learning Systems: Theory and Application. Ed. K. Narenda. New York: Plenum Press.