Explanation or prediction?

<< Click to Display Table of Contents >>

Navigation:  Geocomputational methods and modeling > Geosimulation >

Explanation or prediction?

Just as there are different types of models, each with their own characteristic features, advantages, and disadvantages, there are different ways that a model can be used. Consideration of these allows a clearer understanding of what distinguishes a good or bad model. The following discussion separates the utility of ABM into two broad categories: explanatory (which are often exploratory in nature) and predictive (also referred to as prognostic or descriptive). Readers are also recommended to refer back to the discussion in Section 3.3, Spatial analysis and the PPDAC model, on alternative approaches to problem-solving (the PPDAC methodology), in particular the Planning stage of the process.

The explanatory modeling approach strives to explore theory and generate hypotheses. The primary purpose is not to predict the future behavior of a system, but rather to provide a framework in which past observations can be understood as part of an overall process. Explanatory models generally focus on a specific aspect of a system, placing emphasis on some details of a phenomenon and ignoring others, in the hope that such laboratory explorations will lead to empirically relevant insights. These models purport to be explanatory by stating how reality should be under ideal circumstances, but they do not attempt to reproduce actual systems. The potential drawback of this approach is the lack of analytical methods to empirically evaluate ABM results (see further, Section 8.2.12, Verification and calibration of agent-based models). Furthermore, although explanatory models can provide considerable insights into theory, it is difficult to establish whether the final model is informative about specific real-world systems and scenarios.

The purpose of an agent-based model adopting an explanatory approach could be to program plausible agent behaviors and interactions that, when run as a simulation, produce similar trends and patterns to those observable through the analysis of real-world systems. A model of this nature would produce a ‘candidate explanation’ for the emergence of observed patterns (see Epstein, 1999 for a detailed overview of candidate explanations). The main challenge for such an application, after ensuring agent behaviors and interactions are plausible, is to develop and test alternative models to identify the range of agent representations that might produce given macro-representations.

The predictive modeling approach follows a fundamentally different logic to the explanatory approach. Predictive models are commonly used for extrapolation of trends, evaluation of scenarios, and the prediction of future states. More specifically, changes in initial conditions (e.g. rules governing agent behaviors and interactions, such as information available to agents, constraints upon or incentives for particular agent behavior or movement, etc.), can be used to evaluate the possible effects on the model outcome. Predictive models are designed to mimic real-world systems, and are particularly useful for scenario development and policy decisions. Parker et al. (2003) identify some of the key benefits of predictive modeling: in particular the authors note that by modeling at a fine scale of granularity agent-based models make very good statistical use of data at a fine resolution; in addition, because agent-based models are not constructed to meet a set of equilibrium criteria, it is possible for the model to simulate discontinuous and non-linear phenomena; and finally, by accounting for heterogeneity and inter-dependencies, models can reflect important endogenous feedbacks between processes. However, this leads the discussion to some limitations of predictive models. Any model with positive feedback can create system behavior referred to as path dependence; where a path to a process can be very sensitive to both initial conditions and small variations in stochastic processes. For this reason, predictive modeling of a system containing positive feedbacks can be very challenging. Also, predictive models can be parameterized with too many real-world data. This can lead to an overly-fitted model (i.e. where the model’s calibration is overly constrained to existing data). It may be problematic to generalize such predictive models to a large range of potential outcomes related to the system under analysis, or to analyze alternative systems.

The choice between adopting an explanatory or predictive approach to modeling is not mutually exclusive. This choice depends, in part, on the required precision of the model, which in turn is directly related to the type of information and knowledge that is required. The purpose of a model, including an agent-based model, is not necessarily to capture faithfully all aspects of a system; and this complicates this decision process further. At a fundamental level, an agent-based model can be used solely to enrich understanding of a process that is present within a system through controlled computational experimentation.

The aforementioned decision regarding the model’s purpose is made harder because agent-based models do not fit easily into the classic deductive/inductive approaches to modeling. Scientists use deduction to derive theorems from assumptions, and induction to find patterns in empirical data. An agent-based modeler might use a deductive approach to develop a set of assumptions regarding the behavior and interaction of agents from a body of literature. However, in contrast to classic deduction, the simulated output of agent-based models cannot be used to prove theorems. Thus, a modeler might generate data from several different simulation runs and analyze these results with inductive methods, similar to those employed for analysis of empirical data. However, unlike the conventional process of inductive generalization, the simulated data comes from a rigorously specified set of rules rather than direct measurement of the real-world. The inability to implement ABM in either a purely deductive or inductive manner has led many authors to argue that simulation in general, and ABM in particular, presents a third way of doing science.