Planning support systems (PSS) are information systems that support planning through problem diagnosis, data collection, mining, spatial and temporal analysis, data modeling, visualization, scenario building and projection, and collaborative decision making (Saarloos et al. 2008). PSS involve a wide range of tools that are as diverse as the steps of the planning process itself, from data selection and integration to public participation and negotiation to final policy compromises and implementation monitoring. There are a variety of models integrated within PSS:
Models that are based on geographic information systems (GIS) and spatial statistics. Some examples of GIS-based models used as PSS are: A "What-if system" (Klosterman 1999), the Planning System for Sustainable Development (PSSD) (Hansen 2001), the System for Planning and Research in Towns and Cities for Urban Sustainability (SPARTACUS) (Lautso 2002), and the Population and Land Use Model (PLUM) (WhatIf? 2011).
Models, such as cellular automata and agent-based models (ABMs), that address the spatiotemporal context of the land-use processes through complexity theory and GIS. Linking PSS with complex system models that are capable of handling the complex spatial and temporal dynamics of urban land-use change will increase their utility as tools for urban and land-use planners' decision making and policy evaluation (Torrens 2002, Zellner 2008). Complex system models, based on cellular automata or agent-based approaches, allow the generation of different land-use change scenarios and, therefore, facilitate the assessment and the implications of those scenarios. In PSS, ABMs are models that allow planners to visualize, analyze, and simulate dynamic phenomena emerging from the interaction of individual agents (Saarloos et al. 2008). One of the main advantages of ABMs is their ability to represent the complex and nonlinear interactions among individuals and actors that directly influence land-use change. PSS with ABMs make use of dynamic models that combine spatial processes and human decision making (Parker et al. 2003, Benenson 2004).
Saarloos et al. (2005) have developed a multiagent model for generating alternative land-use plans in which the agents are land-use experts who initiate the development of plan proposals and communicate with each other over time to draw up the proposals incrementally. Li and Liu (2008) have utilized an ABM as a spatial exploratory tool for generating alternative development patterns with sustainable development strategies. Kii and Doi (2005) have examined the effectiveness of policy measures aimed at achieving a compact city form by developing the MALUT agent-based model. Ligmann-Zielinska and Jankowski (2007) have applied the CommunityViz Policy Simulator to generate development scenarios to evaluate ABM's operational use in applied planning settings. Furthermore, Ligmann-Zielinska and Jankowski (2009) have experimented with a utility-based approach and an agent-based model to test different conceptions of risk-explicit decision making. The benefits and drawbacks of the integration of an agent-based model of land-use change and a groundwater flow numerical model have been explored by Zellner and Reeves (2010). They examined the multidimensional effects of land-use patterns and policy implications. Robinson and Brown (2009) have illustrated the use of a GIS-based ABM, called DEED, to produce results that describe the individual and interaction effects of minimum lot-size zoning and land-acquisition strategies on forest cover. Zellner et al. (2009) have employed an agent-based model to explore how underlying microbehaviors affect the payoffs of regional forested space and of local tax revenue obtained by two neighboring municipalities in a hypothetical exurban area. Although all these studies utilize ABMs for urban planning or urban policy evaluation, they are either theoretical (never implemented in real contexts) or they have not been fully integrated into a PSS framework.
Other studies have identified challenges when existing ABMs are used in urban planning. The first challenge relates to simulating agent behavior and interactions. Most ABMs define agent behavior and interactions with rules (Gimblett et al. 1996, Torrens 2006). Rule definition is a challenge in ABM design because:
The large amount of alternatives from which an agent has to choose causes the model runs to be long, producing computational complexity (e.g., endless loops) and consequently modeling error.
The complexity of internal relationships between variables makes the models as black boxes and most of the methods are not designed to handle the complexity.
Additional studies have used random utility theory and multinomial logit models (MNL) (Waddell 2002, Miller et al. 2004, Torrens and Nara 2007). In these models, the variables directly affect the modeling phenomenon and may not account for indirectly interacting variables (in which variable A affects variable B, which then affects the model output). Indirect variable interaction is important because variables in real-world processes do not only affect each other directly. As a result, it is important to include this characteristic in an urban model for it to be realistic. Fuzzy logic has been used for agent decision rules (Graniero and Robinson 2006), but this approach does not provide inference from observed data (Hassan et al. 2010). This is important for analyzing historical trends in urban systems. Neural networks (NN) also have been used in ABM decision rule design (Collins and Jefferson 1992, Gilbert and Terna 2000), but one drawback is that they do not provide information about the relationships between variables (Mas et al. 2004), including how they affect each other and the model output. Genetic and evolutionary algorithms methods also have been used for rule definition to represent human decision making for the ABMs in the land-use change process (Manson 2006). However, the fitness functions of genetic algorithms do not necessarily provide a single solution but a range of solutions and this becomes a disadvantage for agent reasoning and decision making. This limitation makes the agent decision making a complex process because having more than one solution produces longer simulation runs and endless loops.
The second challenge with existing ABMs is the way in which they incorporate dynamic variables and how the system adapts to changing variables. It is difficult to incorporate dynamic variables in the existing models. Their design is not modular and, therefore, they do not allow the easy addition of updated variables. Nonmodularity means variables and their relationships are static and permanent. To add a new variable, the model would have to be recoded and redesigned entirely for on-the-fly changes are not possible. In addition, these models do not handle many dynamic variables in the modeling process. Moreover, there is no change in the relationships between model variables during model iterations as a result of changing conditions.
The third challenge with the existing ABMs is how they deal with uncertainty and limited data about agent behavior. ABMs usually are designed to use historical data for which there may not be adequate quantity and quality. This means the models must rely on assumptions and hypotheses and so uncertainty is introduced into the modeling process. To overcome this problem, a certainty measure can be added to the rules (Batty 2005) by which the rules describe how much a change in the certainty of the inputs will change the certainty of the output. However, uncertainty is not easy to localize; thus, it strongly depends on all the parts of the model.
The Bayesian Network-based Agent System (BNAS) model responds to these aforementioned challenges by integrating Bayesian Networks (BNs) and an agent-based model. The BNAS model has been conceptualized, developed, and validated for modeling land-use change by Kocabas (2008) and Kocabas and Dragicevic (2006). The main objective of this research study is to examine the capabilities of the BNAS model as a policy generator in a PSS framework. By employing BNs in its design, the BNAS model can use backward inference. This means that it is possible to enter a desired land-use pattern into the model and then use it to generate policies that would create that land-use pattern. The resulting simulations can be used as a laboratory for exploring scenarios for urban land-use change. The BNAS-PSS framework was applied to the rapidly growing city of Surrey, British Columbia. Scenarios of land-use change were developed to reflect various urban policies.
BNAS MODEL DESIGN
The BNAS model combines GIS, BNs, and agents to model complex urban land-use change and the spatial relationships between drivers affecting the change. It consists of the following components:
Agent types: The BNAS model includes household agents and commercial firms that settle in a particular urban area. Four different types of agents (three different household agent types and one commercial firm agent type) decide where to live/locate during the simulation periods. Based on socioeconomic characteristics that affect how households evaluate locations, household agents are categorized as high-income (annual income more than C$100,000), middle-income …