Professional Documents
Culture Documents
Research Article
RAJA R. SENGUPTA
Department of Geography, 4438 Faner Hall, Southern Illinois University,
Carbondale, IL 62901-4514, USA; e-mail: rajasen@siu.edu
1. Introduction
Sprague (1980) defined Decision Support Systems (DSS) as computer systems
that are: (i) designed to solve semi- and un-structured problems that upper level
managers often face; (ii) able to combine analytical models with traditional data
storage and retrieval functions; (iii) user-friendly and accessible by decision makers
with minimal computer experience; and (iv) flexible and adaptable to different
decision-making approaches. Extending this definition, Armstrong et al. (1986) used
the term Spatial Decision Support Systems (SDSS) to refer to computer programs
that assist decision-makers generate and evaluate alternative solutions to semi-
structured spatial problems through the integration of analytical models, spatial data
and traditional geoprocessing software (such as GIS). While the potential utility of
SDSS technologies is well recognized, two problems have historically impeded the
development of such systems: (i) geographically referenced data capable of supporting
desired analyses often did not exist and was too expensive to acquire; (ii) the
integration of modelling software and data from disparate sources was beyond the
technological capabilities of many potential users. While the availability of a large
International Journal of Geographical Information Science
ISSN 1365-8816 print/ISSN 1362-3087 online © 2003 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/1365881022000015990
158 R. R. Sengupta and D. A. Bennett
number of spatial datasets on the Internet has helped to alleviate the first problem,
the second remains a significant challenge.
Nijkamp and Scholten (1993) suggest that most decision-makers want an SDSS
to provide them with strategic information without having to learn the technical
details of the system. Building an SDSS, however, often requires the integration of
disparate models and data by individuals possessing considerable technical expertise.
Decision-makers, who are typically infrequent users of GIS software, often lack the
skills necessary to integrate web-accessible analytical models, GIS and spatial data
(Davies and Medyckyj-Scott 1996). The result is a ‘GIS bottleneck’, where they have
to rely on a GIS analyst for generating and evaluating solutions (Armstrong 1994).
Further, even after an analyst has integrated the components for decision-makers,
there may be a ‘conceptual access barrier’ that limits the utility of the newly created
SDSS (Armstrong and Densham 1995). The conceptual access barrier arises from
a lack of understanding of the sequence of map-algebraic operations necessary to
perform spatial analysis, and is complicated further by requiring the user to move
back-and-forth between GIS software and an external analytical model.
An agent-based approach can be used to help ameliorate the issues discussed
above, by emulating the behaviour of GIS analysts. These agents:
(i) assist users locate and retrieve spatial data and analytical models distributed
on the Internet
(ii) automatically transform spatial data for input into analytical models through
the use of GIS software.
The prototype SDSS framework, called the ‘Distributed Intelligent Geographical
Modelling Environment (DIGME)’, described in this paper illustrates the utility of
this approach. DIGME is tested through the creation of an SDSS for the Cache
River watershed of southern Illinois. This SDSS is designed to help evaluate the
ecological and economic impacts of alternative agricultural policies using
data and analytical models located on an intranet at Southern Illinois University
at Carbondale.
2. Background
To meet the objectives of this study we rely on a variety of established and
emerging technologies. System integration using information stored in digital spatial
data repositories provides the overarching motivation for this work. Software agents
are the primary tools that support software development and data integration.
wizards to automate the process of data conversion and provide links to analytical
models from within a GIS software interface. This approach simplifies the user
interface, but forfeits some of the flexibility associated with loosely coupled systems.
Furthermore, users must still work with multiple data formats. In fully coupled
systems, GIS capabilities and spatial models are integrated into a single software
package. User interaction with these systems is relatively straightforward and data
are stored in a common format. There is concern, however, that fully coupled systems
lack the flexibility needed to adapt to changing decision-making environments.
The DIGME framework utilizes software agents to capture the flexibility of
loosely coupled architectures, while maintaining the integrated appearance and ease
of use associated with tight and fully coupled systems.
(iii) respond to changes in their environment; and (iv) exhibit goal-directed behaviour.
Further, some artificial intelligence researchers suggest that humanistic characteristics
such as beliefs, desires, intentions, (Shoham 1993) and even emotions (Maes 1995)
should be a part of agent behaviour. Other attributes that are at times used to
describe agent behaviour include: (i) mobility, where the agents have the ability to
transport themselves at will between various computers on a network (Kotz and
Gray 1999); (ii) veracity, the assumption that agents will not knowingly communicate
false information; (iii) benevolence, the assumption that agents do not have conflicting
goals; and (iv) rationality, the assumption that an agent will act in a manner that it
sees fit in order to solve a problem at hand.
Agents achieve autonomy by sensing activity (through sensors) and modifying
their environment (through effectors) (figure 1). The collective action and reaction
of multiple agents responding to events cause changes to cascade through the system
and modify the state of the environment in which they operate (Russell and
Norvig 1995). Within the framework of spatial modelling, autonomous agents have
been used to: (i) simulate the interactions among users of recreational facilities
(Deadman and Gimblett 1994, Gimblett et al. 2000); (ii) model the ‘actor-institution-
environment’ linkages that cause tropical deforestation and subsequent cultivation
in the southern Yucatan peninsula of Mexico (Manson 2000); and (iii) model animal
populations and behaviour in areas with very low population densities and a large,
diverse, and fragmented terrain (Westervelt and Hopkins 1996). Within the frame-
work of decision-support, autonomous agents have been written to facilitate interop-
eration of different software products (Genesereth and Ketchpel 1994), and promote
the fair representation of stakeholders within a group decision-making scenario
(Edmonds et al. 1994).
While agent technology has been applied to a variety of tasks, the agents utilized
here are a form of ‘autonomous personal agent’ (Maes 1995). In this context, personal
agents cooperate with inexperienced users to initiate and manage complex computer
tasks (Hewitt 1986). Personal agents can therefore: (i) hide the complexity of difficult
tasks from the user; (ii) perform tasks on the users’ behalf; (iii) train/teach users;
(iv) help multiple users share information; (v) monitor events; and (vi) interact with
a changing operating environment on behalf of the user. Personal agents may also
be capable of learning through repeated interaction with the user (see for example
Kautz et al. 1997). An example of this kind of application is an agent that learns to
automatically sort through a user’s electronic mail by observing repeated patterns
in the user’s behaviour (Maes 1995). In DIGME, personal agents collaborate with
users and assist them as they manipulate complex datasets and geoprocessing
software.
3.1. Metadata
Metadata is usually defined as ‘data about the content, quality, condition, and
other characteristics of data’ (Federal Geographic Data Committee 1994, 1). Here
we extend this definition to include information about models and analytical tools.
We contend that model metadata is required to index and categorize the growing
number of spatial models available for download from the Internet. Using model
metadata, it would also be possible to build an indexing structure similar to NSDI
to facilitate the search and retrieval of these resources. Further, we propose that
model metadata consists of the following elements: category (e.g. economic farm
model), name (e.g. GEOLP), input file format requirements, output file format(s),
an assessment of reliability and accuracy, assumptions related to proper use and
application, and a brief description of the model itself. A sample metadata file used
within DIGME for the documentation of a farm-based economic model (GEOLP)
is shown in figure 2. The input and output file format specification is documented
using the Geographic Data Definition Language.
Figure 3 illustrates a metadata file for the documentation of data characteristics.
This file utilizes components of the content standards recommended by the FGDC
for geospatial data (Federal Geographic Data Committee 1994).
language, and GIS software packages. More specifically the following operators were
implemented:
(i) Disjoint, Meet, Equal, Inside, Contains, Covers, CoveredBy, and Overlap,
derived from the 4-intersection model.
(ii) Arithmetic operators, +, −, * and / derived from Map Algebra.
Agent-based modelling environment for spatial decision support 163
(iii) Relational (>, <, >=, <=), equality (equal), negation (not), assignment
(let, setf ), logical (and, or), conditional (if ), and iterative (loop, dolist); these
operators were derived from LISP
(iv) Proximity operators from commercial GIS software (e.g. Buffer).
A hierarchical categorization scheme for geographic objects, called the Geographical
Name System or GNS (e.g. manmade.roads.interstate), was also developed as part
of GDDL, and is described in Sengupta et al. (1996). The GNS can be used to
specify the thematic properties of a dataset. A sample GDDL query requesting soils
data for a specific region (i.e. the Big Creek subwatershed of the Cache River in
southern Illinois) is shown in figure 4.
about spatial datasets available in various digital repositories, and are used in
conjunction with the database blackboard to identify relevant data. The transforma-
tion KSs, along with the transformation blackboard, contain information about
possible data transformations (including transferring data between computers), and
are used by the agents to convert spatial data into the format required by a specific
model. The aspatial blackboard and aspatial KSs perform a function similar to the
transformation blackboard and KSs, but deal with transformations relating to
aspatial databases, such as DbaseB files.
determine the kinds of models that are available in network accessible repositories
and presents a compiled list to the user. If desired, the user can also request more
detailed information about particular models once they are located (e.g. its outputs,
and a general model description). Control is passed onto the model agent when the
user selects a specific model.
Transformation KSs define single functions (i.e. they are atomic). Atomic defini-
tions provide the flexibility needed to adapt to unexpected circumstances. Each
transformation KS is comprised of three elements: a pre-condition clause, a post-
condition clause, and an action function. The pre-condition clause documents the
168 R. R. Sengupta and D. A. Bennett
Figure 11. Pseudo-code for a KS that represents the re-projection of ESRI shapefiles from
latitude-longitude to UTM.
Note that since KSs are atomic and transformation plans are constructed in
incremental fashion, it is possible that a specific action will lead to a computational
dead end. When this happens the system backtracks, instantiates lower rated KSs
and re-initiates the search process.
Further, while some transformation plans (or some portions of it) require the
data transformations to be applied in a particular sequence, there are others for
which the sequence makes no difference. For example, before reprojecting a dataset
it may be necessary to convert the dataset to the format required by a particular
reprojection tool. It may not matter, however, whether a vector dataset is projected
before or after the creation of its topology. Plans where the sequence of some, but
not all operations have to occur in a particular order, are called Partial Order Plans
(POPs). The advantage of using blackboard architectures to develop POPs is two-
fold: while the precondition functions of the various transformation KSs ensure that
there are no incorrectly ordered transformations, the incremental approach in build-
ing the POP ensures that combinatorial complexities that arise as a result of simple
re-orderings of unrelated transformations are avoided.
5. An application of DIGME
To illustrate the utility of the DIGME framework, a working prototype was
created utilizing several computers connected to the Spatial and Environmental
Analysis Laboratory (SEAL) intranet at Southern Illinois University, Carbondale.
A test problem was set-up that required DIGME to implement an SDSS for the
Cache River Watershed in southern Illinois.
The form and structure of this SDSS is patterned after earlier SDSS work in the
region (Sengupta et al. 2000). It is designed to assist decision-makers ascertain the
impact of federal and state conservation policies on non-point source pollution and
the economy. This traditionally built SDSS is comprised of two analytical modelling
software packages, GEOLP and AGNPS, tightly coupled with a GIS software
Agent-based modelling environment for spatial decision support 171
Figure 12. Data transformations suggested by blackboard architecture for GEOLP data
requirements.
wide web. Three technical issues may, however, impact the ready adoption of these
kinds of Internet-based solutions: (i) availability of bandwidth; (ii) accessibility to
computing resources; and (iii) scalability.
The availability of adequate bandwidth is critical to the success of an Internet-
based DIGME as data transfer among computers can take significant amounts of
time. While this may not be a major problem for servers and desktop computers
connected to the Internet via broadband technologies (i.e. high-speed connections),
it can hamper the performance of DIGME for devices that rely on slower wireless
or telephone network connections (Kotz and Gray 1999). A simple solution to this
problem is to monitor and manage the size of files being transferred among machines.
In the existing version of DIGME, this is achieved by assigning ratings to the file
transfer process that are inversely related to file size. Another solution would be to
use Remote Method Invocation (RMI) and Object Request Broker (ORB) technolo-
gies to allow temporary file and directory sharing among computers (Wade et al.
1997). This would also reduce the computing resources required by the end user by
capturing the processing and storage power of several computers.
Issues associated with access to computing resources include security permissions,
adequate disk space on servers for remote processing, and alternative locations for
the retrieval of data (mirror sites) in case of network failure. Implicit in the creation
of DIGME is the assumption that mobile agents will have access to adequate
computing resources (such as disk space and CPU time for executing models) on
remote servers. While issues of adequate disk space and CPU time can be viewed
simply as resource constraints, the real challenge is to provide DIGME agents with
full access to these computing resources without being hampered by security designed
to prevent attacks by malicious agents (Kotz and Gray 1999). Research designed
to address the security problem while maintaining agent operability in an open
environment is being conducted (Vigna 1998).
174 R. R. Sengupta and D. A. Bennett
create an SDSS on their own by utilizing model and data sources downloaded from
Internet-based repositories. According to Nijkamp and Scholten (1993), most upper
level managers and concerned citizens fall into this category. One concern is that
DIGME makes it easy for uninformed decision-makers to run models without
understanding the assumptions and errors that modelling entails. However, this
danger is not limited to DIGME alone. Watson and Wadsworth (1996) mention
that users of the NELUP DSS designed to model rural land use change often did
not question the results generated by the models. A potential workaround is to
present model assumptions to users and to warn users when violations of these
assumptions occur.
Despite its shortcomings, DIGME is a promising technology that utilizes agents
and metadata to provide decision-makers with greater choice and flexibility in
integrating models and data into decision support technologies. In this sense,
DIGME acts as a ‘DSS generator’ originally envisioned by Sprague (1980).
Acknowledgments
The authors would like to thank Dr Norman Carver of the Computer Science
department at Southern Illinois University for ideas and discussions related to
blackboard system design and implementation. Financial support for this research
was provided by grants from the Illinois Council for Food and Agriculture, and the
Pontikes Center for Management of Information, Southern Illinois University,
Carbondale.
KS Action Precondition
RelJnA-KS Triggers the search for aspatial The attribute names in the
datasets with a posting on the available dataset must not
aspatial blackboard match those desired by the
model. In addition, the spatial
database that prompts
RelJnA-KS to start searching
for aspatial data should be in
ESRI shape file format on
computer3
RelJnB-KS Performs the relational join The preconditions are the same
with the dataset that started as for RelJnA-KS. In addition,
RelJnA’s search for aspatial a solution must be available on
data the aspatial blackboard
References
A, M. P., 1994, Requirements for the development of GIS-based group decision
support systems. Journal of the American Society for Information Science, 45, 669–677.
A, M. P., D, P. J., and R, G., 1986, Architecture for a microcomputer
Based Spatial Decision Support System. In Proceedings of the 2nd International
Symposium on Spatial Data Handling (Williamsville: International Geographical
Union), pp. 120–131.
A, M. P., and D, P. J., 1995, GIS and group decision-making: a look at the
dark side. In Proceedings of GIS/L IS ’95 (Washington, DC: American Congress on
Surveying and Mapping), pp. 11–19.
A, M. P., R, G., H, R., D, B. T., L, P., D, S., and
D, P. J., 1991, Decision Support for Regionalization: A Spatial Decision
Support System for Regionalizing Service Delivery Systems. Computers, Environment
and Urban Systems, 15, 37–53.
B, D., 1997, A framework for the integration of geographic information systems and
modelbase management. International Journal of Geographical Information Science,
11, 337–357.
B, B. P., and G, M. F., 1996, The Alexandria digital library project:
distributed library services for spatially referenced data. In Proceedings of GIS/L IS ’96
(Washington, DC: American Congress on Surveying and Mapping), pp.76–84.
C, N., and L, V., 1994, The evolution of blackboard control. Expert Systems with
Applications, 7, 1–30.
C, D. D., 1991, Blackboard systems. AI Expert, 6, 40–47.
D, C., and M-S, D., 1996, GIS users observed. International Journal of
Geographical Information Systems, 10, 363–384.
D, P., and G, H. R., 1994, The role of goal-oriented autonomous agents in
modelling people-environment interactions in forest recreation. Mathematical and
Computer Modelling, 20, 121–131.
E, E. A., C, L., J, R., and S, B., 1994, Support for collaborative design:
agents and emergence. Communications of the ACM, 37, 41–47.
E, M. D., and F, R. D., 1991, Point set topological spatial relations.
International Journal of Geographical Information Systems, 5, 160–174.
E, L., H-R, F., L, V., and R, D. R., 1990, The hearsay-II speech
understanding system: Integrating knowledge to resolve uncertainty. Computing
Surveys, 12, 213–253.
F G D C, 1994, Content Standards for digital spatial metadata
(8 June draft). http://www.fgdc.gov/metadata/contstan.html
G, M. R., and K, S. P., 1994, Software agents. Communications of the ACM,
37, 48–53.
G, H. R., R, C., and D, T. C., 2000, Intelligent agent modelling for
simulating and evaluating river trip scheduling scenarios for the Grand Canyon
National Park. In Integrating GIS and Agent based modelling techniques for
Understanding Social and Ecological Processes, edited by H. R. Gimblett (Oxford:
Oxford University Press), pp. 245–275.
G, M. F., 1992, Integrating GIS and spatial data analysis: problems and possibilities.
International Journal of Geographical Information Systems, 6, 407–423.
H, C., 1986, Offices are open systems. ACM T ransactions on OYce Information Systems,
4, 271–287.
K, H., S, B., C, M., and K, S., 1997, An experiment in the design of
software agents. In Readings in Agents, edited by M. N. Huhns and M. P. Singh (San
Francisco: Morgan Kaufmann Publishers), pp. 125–130.
K, D., and G, R. S., 1999, Mobile agents and the future of the Internet. ACM Operating
Systems Review, 33, 7–13.
L, S. E., 1995, Blackboard-Based Integration of Legacy and Reusable Software
Systems. NSF SBIR Final Report Contract No. 94-45, National Science Foundation,
Washington, D.C.
M, P., 1995, Intelligent software. Scientific American, 273, 84–86.
180 Agent-based modelling environment for spatial decision support