 |
|
ISSN 1320-0682 | ||||
| Volume 3 | April 1996 | ||||
Randy Gimblett, Bohdan Durnota & Bob Itami
...Gimblett
School of Renewable Natural Resources, University of Arizona, Tucson AZ 85721,
USA
...Durnota
Faculty of Computing and Information Technology, Monash University, PO Box
278, Caulfield East 3145, AUSTRALIA
...Itami
Digital Land Systems Research, PO Box 4191, Parkville 3052, AUSTRALIA
Email:gimblett@nexus.srnr.arizona.edu.au, bdurnota@fcit.monash.edu.au & dlsr@werple.mira.net.au
Increasing pressures by recreators on public lands has forced U.S. federal agencies such as the Forest Service to re-evaluate forest management activities and adopt new practices for assessing recreation use. Because natural ecosystems, including the human dimension of such systems, are extremely complex, integrated systems, models capable of linking and spanning multiple production processes, and geographic and temporal scales, are needed to support forest management decisions. The goal of the research reported in this paper is to develop a new form of intelligent decision support and simulation system (IDSS) which uses autonomous agents to assist natural resource managers in assessing and managing dynamic recreation behavior, social interactions and resulting conflicts in wilderness settings. We present a framework for modeling recreation conflicts between recreation use groups in the redrock country of Arizona, U.S.A. The power of GIS is utilised for accurately representing complex dynamic landscapes, and together with advanced vision detection and assessment capabilties, is used for simulating conflicting recreational activities in these landscapes. We describe linkages between the dMARS Distributed Multi-agent Reasoning System, the Swarm Multi-agent Simulation System and a GIS system to develop goal-oriented autonomous recreation agents capable of reasoning and decision-making in their quest to seek and derive satisfactory recreational experiences, while minimising recreational conflicts in crowded backcountry environments. We briefly describe each of the components of the framework and outline our approach for calibrating our agents against empirical recreation data collected in Sedona, Arizona for this study.
Increasing pressures by recreators on public lands is forcing U.S. federal agencies such as the Forest Service to re-evaluate forest management activities and seek new practices for managing recreation use in wilderness. Because natural ecosystems, including the human dimension of such systems, are extremely complex, integrated systems, models capable of linking and spanning multiple production processes, and geographic and temporal scales, are desperately needed to support forest management decisions. What theoretical research and development that has been undertaken in the field of recreation management over the last decade has tended to focus on the issue of recreation-carrying capacity. This concept, which is derived from range management, has had limited success in recreation management. Part of the problem has been that research in recreation management in general has been severely criticised for its lack of theory and over-dependence on survey techniques for exploring human attitudes and behaviour in outdoor environments. This has left recreation resource managers, who undergo constant staff and budget reductions, with limited new theories, information and/or methods and tools for assessing and resolving recreation use and associated conflict issues in landscapes which are suffering from extreme human pressure and recreation use.
A recent policy directive from the U.S. Forest Service to all forest service units has been to adopt Ecosystem Management into all forest level planning. This directive, which explicitly calls for the incorporation of the human dimension into ecosystem management, has set the stage for some serious re-evaluation of management decision-making, and has led to the consideration of human values and use in public lands. Within this human dimension of ecosystem management, recreation and amenity uses of forest lands and the associated benefits of those uses, constitute an important component of management decisions. It would be useful for managers incorporating the human dimension of ecosystem management into forest planning if integrated system decision support models incorporated explicit functional relationships between manageable characteristics of forest lands and the recreational uses of those lands. Such explicit relationships would provide clear linkages between management activities for recreational and non-recreational uses of forest lands, and between management activities for recreational uses. These relationships can then be incorporated into decision models to assist managers in recreation and amenity planning.
While our work is in its infancy, it aims to create a new-form intelligent decision support system (IDSS) to assist natural resource managers to assess and manage dynamic recreation behaviour, social interactions, and resulting conflicts in wilderness settings. The focus of this research is to utilise simulation techniques to explore the complex interactions between recreators and the environment, and interactions between recreators, as a means to improving the foundations of recreation theory. Specifically, the research aims at developing a simulation model which incorporates GIS technology with artificial autonomous agent models of recreators to simulate recreation behaviour in real environments. Geographic Information Systems (GIS) are being extensively used in many forests to aid in making more informed decisions about resource allocation. While much of the focus of GIS-models has been on the biophysical environment, the wider human dimension has been largely neglected. One strategy for incorporating this human dimension is to use intelligent agents in multi-agent simulations, and look for emergent properties in the interaction of various multi-agent scenarios. Since GIS aids in constructing precise, geo-referenced databases of the real world, agents relevant to modelling the human dimension need to be calibrated against human behaviour in such worlds.
In the literature which touches upon multi-agent systems, such as distributed artificial intelligence, there are not many papers which deal with simulating man-(natural) environment linkages. An example includes the Phoenix system [13] and [36] which has been constructed to explore forest fire-fighting (in Yellowstone National Park), and uses autonomous fire-fighting agents. Another example is the multi-agent simulation of fisherfolk on the Niger inland delta [9]. Other relevant work includes that being done by archaeologists [17], [16], [38], [51] and [48].
The paper is organised as follows. The section on GIS, The Biophysical Environment Human Behavioural Modelling provides a perspective on modelling the natural environment and human behaviour. The section on Bounded Rationality Calibrated Agents emphasises the task of designing human-like artificial agents and their calibration against empirical data. Finally, the section on The Project Framework briefly describes how our research is addressing these problems in a natural resource context.
While much research has been done to understand and model biophysical interactions in the environment using GIS [22], recent work has begun to focus on incorporating developments in Global Positioning Systems (GPS) in acquiring much more accurate data. This data can be used to improve the accuracy and expand the range of applications and types of simulations possible. With advanced modelling and decision-making needs, intelligent systems, object-oriented programming techniques and parallel processing for such simulations has come to the forefront of research [21], [42] and [6]. While much work has been done in GIS and advanced spatial modelling, very few to date have attempted to incorporate social science data into spatially-explicit models of human behavioural phenomena.
One set of applications that so far h as not fully explored the use of GIS to accurately represent the biophysical environment is the field of artificial intelligence (the Phoenix project [13] and [36]) being one of the few exceptions. Automated systems research such as that of robotics and artificial life have been slow to embrace the advantages of using geo-referenced spatial data. Researchers have attempted to link object-oriented techniques with simulated two-dimensional lattice environments [47], [27] and [6], while others such as those exploring emergent behaviour in the field of artificial life have used perhaps even more contrived environments. While the biophysical environment was certainly not the focus of these research efforts, it could provide some added advantages of complexity, as found in real environments, to the simulations.
In addition to the work outlined above, the recent explosion in the use of intelligent agent simulations have come to the forefront of Artificial Intelligence (AI) research. This research includes both intentional agents, whose behaviour is driven by the constructs of folk psychology; and reactive agents, whose behaviour is usually very simple but whose interactions amongst a community of such agents can give rise to interesting emergent properties. Both types of agents interact and react with their environments; hence, they are active objects, and they have at least some partially determined set of goals.
The reactive agent school of thought professes that agent behaviour is viewed as an intensive interaction and emergent property between that of the agent and its dynamic environment. The environment is not only taken into account dynamically, but its characteristics are exploited to serve the functioning of the agent. A complex agent has complex goals, but behaviour to meet the goals develops or emerges as the agent learns more about its surroundings. Recent examples of tools and testbeds developed around this work include GENSIM ( [2]), Swarm ( [24], [35] and [46]) and Echo ( [30], [19] and [18]). While all of these simulations have been successfully implemented in contrived environments, real biophysical environments have not yet been explored. The only work known to the authors that attempts to build direct linkages between such (reactive) multi-agent simulation models and GIS is by [10] (using GRASS and STELLA modelling software), and [50] (using GRASS and object-oriented techniques into a Spatial Modelling Environment). [49] has mentioned recent work which is the first attempt to link Swarm and GRASS for natural resource agent modelling.
On the intentional side, there are a number of multi-agent systems such as the dMARS Distributed Multi-Agent Reasoning System [40] and [15], but apart from the Phoenix system [13] and [36], the authors do not know of any that have used GIS as a foundation. In fact, only a few examples are known in which a form of intentional agent architecture has been applied to study natural resource management problems [9], [1], [42] and [14]. Nevertheless, there is much current work in this field of direct relevance for the modelling of man-environment interactions. For example, [7] explore how belief systems may be simulated in agents.
Our contention in this paper is that modelling the human dimension in natural resource management can appropriately utilise both the above conceptions of agency. On the one hand, we wish our artificial human-like agents to be defined in intentional terms. However, how well one is able to calibrate such intentional agents using empirical data is an important and interesting exercise. However, on the other hand, we wish to study the emergent properties of collections of such agents, and the viewpoint and tools of the complex adaptive systems community are appropriate here.
One area of social research which has begun to use agents in simulating complex social interactions is that of economics [25] and [39]. Artificial economic agents are being designed to explore the complex interactions in their behaviour among themselves and with their economic environment [26], [3], [4] and [45]. For example, [32] have explored the dynamics of a financial market in which different agents utilise different trading strategies (see also [43]), whilst [11] have explored the effects of strategic interactions in oligopolistic markets; that is, interactions between agents in which other agents' behaviour is taken into account in arriving at a decision between agents. One reason for the current interest in these artificial agents is that they allow us to go beyond the limited assumptions of economic rationality; to incorporate various limitations on time, resources, knowledge and other factors in the decision-making processes of agents - that is, bounded rationality - and of learning and adaptation [37], [8] and [12].
Since it is likely that different decision-making contexts may result in different actions, if we attempt to design a human-like artificial agent, we may end up with a multitude of different algorithms. To bound this modelling effort, Arthur [3] and [4] has suggested that we calibrate the behaviour of these artificial agents against real data; hence the name of such agents - calibrated agents. The agents should then be able to be applied in new situations with a greater degree of confidence.
What emerges out of this discussion is a number of interesting developments in both the AI arena as well as the problem of accurately representing the three- and four-dimensional spatial environment. While tools are being developed to more accurately capture, assess and represent dynamically functioning ecosystems, object-oriented agent platforms are being developed with task-driven, goal-oriented behaviours, which in many cases result in emergent phenomena. While this research is exciting, it requires much further development on both fronts if they are to be successfully integrated. Agents need to be capable of accurately reading and interpreting biophysical environments based on a well-grounded cognitive portfolio. Agents must have emergent behavioural qualities, and be able to evaluate, respond and adapt to complex, dynamic environments. Above all, for decision-making in our practical forest management applications, human-like agent-based simulations must be realistic, accurate, and grounded in reality like Arthur's calibrated agents through the use of empirical data so that they can be applied in new management situations with greater degrees of confidence.
For the modelling requirements of this project, our intelligent decision support framework combines:
Figure 1 represents the overall framework for our modelling system.
Figure 1: Framework for modelling recreation behaviour
Sage Professional ( [28] and [41]), a state-of-the-art GIS system, is being used to initially build a spatial database from aerial photos, topographic and other available spatial data. Figure 2 shows a terrain model associated with the Sedona study site.
Figure 2: Terrain at the Sedona study site
Sage has been recently converted into object-oriented C++ to make the integration between Swarm and dMARS more transparent. Direct linkages are being made from Swarm/dMARS to the GIS database. This enables physical and positional data to be read directly from GIS databases, alleviating the modelling of recreation suitability, and the importation of a lattice into Swarm. This direct GIS linkage provides more flexible autonomous behaviour for the goal-oriented agents seeking an appropriate location in the landscape.
One of the most powerful spatial functions in a GIS system that is frequently used by visual resource managers are those that perform visibility analysis. Within current GIS systems, viewshed analysis provides the modeller with a computed surface that assigns ordinal values to cells which can and cannot be seen from a particular location, the frequency of which objects can be seen, and at what distance [29]. While this type of analysis provides exceptional visibility measures of the landscape, it assigns values to the cells being seen, rather than reporting back information to the cell being viewed from. In addition, there is no associative meaning or cognitive framework to identify what the content of the landscape means in terms of the data collected from the GIS and human behaviour and value systems. [34] suggest that researchers in environmental perception have concluded that personal experience of landscape can be divided into four categories:
Likewise, [5] outline the importance of exploring the relationships between physiographical, feature and cognitive components of the GIS environment. For our research, this is extremely important. Not only do any modifications of viewshed algorithms need information to be communicated back to the viewer, but this information must be put into a conceptual framework in terms of human behaviour if our goal-directed agents are to move around in the GIS environment.
While many agent systems have provided limited vision systems , our work extends those capabilities. Modifications made to the new Sage C++ object code provide our agents with full vision capabilities in GIS environments. The algorithm calculates the contribution of each visible entity to the viewer's field of vision, measured in square degrees. This information is then overlayed onto the physiographic classification of the landscape. A table consisting of the visual magnitude of each physiographic unit may then be reported, along with other quantitative summaries of the data such as mean distance of the unit from the viewer, mean distance to important landscape features, and position within the visual field. In addition, the vision system is capable of detecting any other agents in its visual field and calculating the relative distance to them.
Linkages between the physiographic classification of the landscape and the inter-visibility table is sent to the recreator agent. Beliefs, attitudes, desires and intentions assist in the translation of this information through a planning scheme and appropriate actions are passed back to the simulation for appropriate movement.
Our agent designs are based on the BDI agent architecture ( [40], [20] and [33]), which is implemented in the dMARS system. This architecture is explained by [33] thus:
consists of a database containing current beliefs or facts about the world; a set of current desires (or goals) to be realised; a set of plans describing how certain sequences of actions and tests may be performed to achieve given goals or to react to particular situations; and an intention structure containing those plans that have been chosen for [eventual] execution.
The system uses an interpreter to manipulate plans which are appropriate, based on the agent's belief system. At any point in time, certain goals are established and certain events occur that alter the beliefs held by agents.
In our initial agent designs, we have been developing several different sets of initial beliefs and plans. Each of our recreator agents has its own set of beliefs about the world, goals and strategies (plans) for obtaining those goals, and can respond to other agents in their immediate vicinity and within their visual field. These intentional structures have been extracted from the empirical field surveys undertaken on site. In addition to these basic behavioural beliefs and actions, we are attempting to derive meaningful cognitive-oriented plans that translate the physiographic data reported from the vision system into value-based actions. A structured set of plans is being developed to respond to a certain set of physiographic conditions, triggering certain movements and behavioural responses in the environment.
The use of simulation is an important ingredient in the design of our IDSS. It provides us with the capability of deriving statistically meaningful conclusions for our multi-agent world. This is crucial for the application of our system in providing advice to natural resource managers. The methodological role of (controlled) experimentation in the design of agent architectures is also increasingly becoming more recognised in the wider AI community [23].
To support such multi-agent simulations, the Swarm system is being used, and is considered to be a core component within the overall IDSS. Software linkages are being designed and constructed to combine components of dMARS with Swarm. Such integration enables us to take advantage of the graphic user interface and simulation environment of Swarm with the design and construction of goal-oriented cognitive agents in dMARS.
While much work has been done in designing artificial agents to perform specific tasks, much work still needs to be done to ensure these agents perform their assigned tasks according to a set of realistic rules. Since it is likely that different decision-making contexts may result in different actions, the design of an artificial agent that behaves like humans in a given context or environment must be bounded or grounded with real data. Simple algorithms calibrated against experimental data have been demonstrated to be usable as the basis for agent behaviour in larger theoretical models to help answer questions about resource allocation in a more general setting.
In this research, physiological and psychological data has been collected and used to calibrate our agents in a simulated GIS world. In order to calibrate our agents to perform tasks to solve real world problems (such as seeking optimal recreation management outcomes), an onsite survey/interview experience sampling methodology was devised. It recorded pertinent behaviour data over an eight-month period of the recreator groups that frequent the Broken Arrow Canyon of the Coconino National Forest in Arizona, an area that had lead to management concerns.
The types of data that were solicited from recreators included:
This data is being used as an empirical measure of goals and intentions of recreators. Detractors and the inability to obtain desired benefits together are used as empirical evidence of goal interference and conflicts, and would imply an inability to obtain desired recreational experiences and of unsatisfactory outcomes. Many of these concepts have been explored by others and are summarised in [44]. All these information sources are used in formulating the intentions and the resulting plans of actions for the artificial recreator agents in the simulations. In addition, the survey captured time in/out, the estimated duration of a visit, and employed Global Positioning Systems (GPS) tracking of trails used by recreator groups in the canyon. This information is being used as an empirical measure of when, how fast, and how many of the different recreator groups frequent the canyon lands at any time over an eight month period.
Three classes of recreator agents have been extracted from the experience sampling methodology (mountain bikers, hikers and pink jeep outfitter tours) and aggregated into activity groups in which individuals have profiles representing the amount of solitude being sought, group responses, group dynamics, goals and beneficial desires, norms, etc. This analysis has assisted in the understanding the diversity of recreational types, settings that are crucial for satisfying recreator goals and desires, and the influence of recreator density and encounters on goal satisfaction. For each class of human agent, behavioural rules are being derived that will provide the data necessary to program artificial agents to perform and react in similar fashion.
The lack of empirical data in this study to calibrate the visual system has prevented us from comparing agent-generated views against human subjects in psychological experiments to identify suitable cognitive criteria used by individuals for evaluating quality of landscapes. For this study, some assumptions have been made between physiographical features and cognitive components of the GIS environment from a decade of research work undertaken in this area [31].
This paper has presented the basic multi-agent framework for our intelligent decision support system. By merging the simulation capabilities of Swarm with advanced cognitive reasoning capabilities of dMARS, a new form of sophisticated decision support system is being created. More importantly, we feel this work is extending the capabilities of the research into multi-agent systems by providing direct connections to accurate, GIS-based environments, capable of reflective real world dynamics, enhanced visibility capabilities that interact directly with this GIS data, and a method for calibrating artificial agents against human behavioural characteristics. These calibrated autonomous agents, when sent traversing the landscape, seek recreational benefits and assess conflicts while satisfying primary and secondary goals. These agents reveal new ways of applying artificial agents in solving real world, natural resource-based problems. Even the preliminary results of this ongoing work clearly reveal the promise of using calibrated, multi-GIS-based agents as a practical method for assessing recreation conflicts in complex landscapes, and for assisting field managers in making more informed decisions for managing recreation use in wilderness settings.
We wish to thank the U.S. Forest Service for providing the opportunity to conduct field work in the Sedona area of Arizona, and the Australian Artificial Intelligence Institute (AAII) for allowing us to use dMARS.
Spatially-Explicit Autonomous Agents For Modelling Recreation Use In Complex Wilderness Landscapes
This document was generated using the LaTeX2HTML translator Version 96.1 (Feb 5, 1996) Copyright © 1993, 1994, 1995, 1996, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
The command line arguments were:
latex2html gimetal3.
The translation was initiated by Pam Milliken on Wed Jan 8 14:29:11 EST 1997Wed Jan 8 14:29:11 EST 1997