• Tidak ada hasil yang ditemukan

Agent based modelling

Glossary of Terms

Chapter 3: Methods Review

3.3 Complexity modelling

3.3.1 Agent based modelling

Epstein and Axtell (1996) in their classic book Growing Artificial Societies: Social Science from the Bottom Up, describe the use of social simulations using Agent Based Modelling (ABM), for

78 analysing social structures and group behaviours. This is a ‘bottom up’ paradigm in that instead of describing aspects of the world using macro-properties, Epstein and Axtell consider the interactions and actions of individuals. In this way, they let the macro-properties emerge rather than being explicitly modelled. The primary question that Epstein and Axtell ask is: ‘How does the heterogeneous micro-world of individual behaviours generate the global macroscopic regularities of the society?’ This is of course the global-local interaction referred to in the previous section; one of the key strengths of the approach.

In a general sense, objectives of the applications of ABM can be broadly categorised into the following categories: prediction, policy analysis, inductive reasoning about system properties, co-learning and collective decision making, developing a systems understanding and knowledge management.

• Prediction (Samanidou et al. 2007): using models to predict some dependent variable into the future so that typically financial gains or competitive advantage can be realised in some manner;

• Policy analysis (Heckbert and Smajgl 2005): using models to help policy makers understand implications of policy to make these more efficient and fair. A particular advantage of ABM in this area is that the method accounts for more realistic behaviour as well as understanding the impacts on a heterogeneous population;

• Inductive reasoning about system properties (Rixon et al. 2007a): using models to understand systems where sufficient micro-level measurement is difficult, and using the simulation model as a virtual laboratory exploring possible realities. This is particularly powerful when it is possible to re-design the actual reality to match the behaviour that is modelled, and hence creating a more predictable and presumably more efficient system;

• Co-Learning and collective decision making (Barreteau 2003): helping groups of stakeholders to understand complex interactions with each other and their environment, typically in order to identify collaborative solutions to resource dilemmas;

79

• Developing a system understanding (Tillman et al. 1999): using models as a tool in the scientific process, exploring and evaluating theories; often in the social science area where such process tools have previously been somehow lacking;

• Knowledge management (Dignum 2006): using models to embed knowledge and information in an accessible manner to multiple stakeholders.

Gilbert and Troitzsch have written a very useful reference book to guide the undertaking of social science through computer simulation:

With complex models, especially if the specification is nonlinear, such analytic1 reasoning can be very difficult or impossible. In these cases, simulation is often the only way. Simulation means ‘running’ the model forward through (simulated) time and watching what happens. Whether one uses an analytical technique or simulation, the initial conditions, that is, the state in which the model starts, are always important.

Often, the dynamics is very different depending on the precise initial conditions used (1999: 16).

In particular, the ABM methodology is good at modelling emergent phenomenon. Emergent phenomenon is described as follows:

Emergence occurs when interactions among objects at one level give rise to different types of objects at another level. More precisely, a phenomenon is emergent if it requires new categories to describe it which are not required to describe the behaviour of the underlying components. An individual atom has no temperature but a collection of them does (Gilbert and Troitzsch 1999: 11).

Hence, this links it back to an essentially thermo-dynamic view of the world, which is signified by interactions between different scales. Or as described by metaphor: in an ABM model of a society, the individual is the atom and their individual characteristics and interactions give rise to macro-properties like the temperature and entropy of thermodynamics.

1This refers to models for predictions (of macro-properties) based on reasoning, such as logic or mathematics.

80 It is however acknowledged humans are more unpredictable than atoms. In fact, attempts to understand our cognitive abilities using the same cognitive abilities are inherently circular (Maturana and Varela 1980). Socio-cultural worlds are strongly contextual and individual people operate in multiple such worlds (Maturana and Varela 1980). Serious attempts in neo- classical economics to define human behaviour have been met with criticism from different angles; relating to the use of heuristics (Gigerenzer and Todd 1999); the evaluation of rewards (Gintis 2000); and dealing with uncertainty (Tversky and Kahneman 1981). In parallel with these more modern theories of behaviour, there is an entire area of scientific endeavour in neuroscience that is still very much exploring surprisingly complex inner workings of human brains, including those relating to feelings and emotions (Johnson 2004). Suffice to say that atomistic representations can never fully capture the complex inner workings of humans, but that it has the capacity to embed at least those representations of human behaviour that can be formalised in mathematical or computational language.

Despite its limitations, atomistic computer metaphors of humans have been surprisingly successful. As an example, Lansing and Miller (2003) describe how social structures and practices have emerged in Bali rice farming communities from only local dynamics and interactions (Lansing and Miller 2003). Indeed, the approach (as described by for example Ferber 1999) has proven to be successful in explaining the emergence of many phenomena based on simple and intuitive local rules of interaction and has even been able to explain widely diverse systems that previous research paradigms has failed to explain, such as racial segregation in cities (Schelling 1978), bird formations (Reynolds 1986), distribution of wealth in societies (Epstein and Axtell 1996), spread of HIV (Wilensky 1998), dynamics of electricity markets (North et al. 2002), or village formations of the Anasazi Indians (Diamond 2002).

However, as previously hinted in the previous section, there is also appropriate criticism of this approach (as described by Lansing 2002 and Richardson 2003). Lansing writes:

One does not need to be a modeller to know that it is possible to ‘grow’ nearly anything in silico without necessarily learning anything about the real world (2002: 2).

81 Whilst Lansing argues for caution in application of ABMs and a more scientific critique, he also agrees that models are often particularly useful for understanding coupled human and natural systems and that they support inter-disciplinary inquiry.

On a similar line of scientific critique of the ABM approach, Richardson (2002) points out that there are infinitely many ways of representing the same phenomenon (equifinality), and that it is not sufficient validation of a model that an actual phenomenon can be predicted in a computer simulation. Consequently it needs to be recognised that while ABM opens up new possibilities for modelling human interactions, the approach also has limitations. It is important to realise that without backing up this tool with rigorous processes, facts and data, it provides little more than a computer game. Validation of computer simulation models has proven to be a difficult task, but one alternative approach is being done as part of the Companion Modelling framework (Barreteau et al. 2003) where the validation is through an evaluation with actual stakeholders, supporting and generating a socially accepted model of reality. Naturally, this type of validation is limited to situations where the stakeholders are available for discussion and evaluation.

Technically, ABM is based on a foundation of object-oriented programming which is usually, but not always, embedded into an Integrated Development Environment (IDE) or a development platform (Rixon et al. 2005). The conceptual representation of a system is referred to as a Multi Agent System (MAS) and the simulation instance is referred to as an Agent Based Model (ABM) and Perez and Batten provide a definition of the components of a MAS:

• An environment (E), often possessing explicit metrics;

• A set of passive objects (O), eventually created, destroyed or modified by the agents;

• A set of active agents (A), Agents are autonomous and active objects of the system;

• A set of relationships (R), linking objects and/or agents together; and

• A set of operators (Op), allowing agents to perceive, create, use or modify objects (2006:

27-28)

82 The structure of these components, and their place within the MAS, is set up in the modelling design process, where the design output is a conceptual model described using Unified Modelling Diagrams (UML) (Bousquet 2004), which has the following components:

The class diagram describes the entities of the modelled system (classes) with their internal characteristics (attributes and methods) and external links with other classes. It corresponds to the casting of the model. The sequence diagram describes the successive actions conducted independently by different classes or interactions between several classes. It corresponds to the storyboard of the model. The activity diagram describes the intimate actions embedded into a given method. The exhaustive list of all the activity diagrams corresponds to the script of the model (Perez et al. 2006: 206).

Figure 3-2 shows a UML class diagram from the model by Ducrot and colleagues (2004) which focuses on the management of the peri-urban interface between land use and water resources in a case study in Sao Paolo, Brazil. The grey boxes show classes (i.e. objects), and their attributes, within the software system, and arrows show the relationship between classes. For example, each plot belongs to an Object Location, and consequently there is an arrow from the Plot class to the ObjectLocation class. Similarly, an ObjectLocation is a special case of a Spatial Entity and consequently, there is a different type of arrow showing this relationship.

In terms of the design of an ABM, Hare and Deadman (2004) have reviewed eleven case studies in Environmental Modelling contexts and have identified key design features and set up a taxonomy of Agent Based Models with the key features being described below.

The first issue relates to how social and environmental models are coupled. In other words, this relates to whether a particular model has a spatially explicit or spatially non-explicit representation. If spatial features are not important for the problem, then a spatially non-explicit model, as exemplified by the model by Janssen and Carpenter (1999) and the SHADOC model by Barreteau and Bousquet (2000) is used. If the spatial environment is important, then a spatially explicit model is used, such as the CATCHSCAPE model by Becu and colleagues (2002).

83 Figure 3-2: Example UML Class Diagram

Source: Ducrot and colleagues (2004)

Similarly, Figure 3-3 shows a UML sequence diagram, showing the sequence of events in the model, and how the model steps through a number of activities carried out by the various classes.

The second issue relates to how decisions are made at the micro-level by agents and this is related to what was described in the section on knowledge engineering. The types of models range from the use of the rational actor assumption where an agent maximises a utility function as per the classical homo economicus definition of classical economics which has been heavily criticised by Gintis (2000) among others; to the use of cognitive models such as Consumat (Jager et al. 1999; Janssen and Jager 1999) or Belief-Desire-Intention (Brazier et al. 2002) capturing concepts mainly from sociology and psychology; to the use of sophisticated knowledge based rule inference such as applied in the SHADOC model (Barreteau and

84 Bousquet 2000) or the ATOLLSCAPE model (Dray et al. 2006a, 2006b). Lansing and Kremerin (1994) describe perhaps another type of micro-decision making model based on simple single behaviour agents using imitation based on social interaction as the driving force for adaptation and change.

Figure 3-3: Example UML Sequence diagram Source: Ducrot and colleagues (2004)

85 Thirdly, another important aspect of the design of an ABM is the level of social interaction between agents, ranging from no direct social interaction but interaction via the environment, as in the CATCHSCAPE model by Becu and colleagues (2002); to social adaptation through imitation (Lansing and Kremerin 1994); and to networks of communicating agents such as in the SHADOC model by Barreteau and Bousquet (2000).

Fourthly, Hare and Deadman (2004) also categorise ABMs on the basis of the mechanisms for adaptation; i.e. how do agents respond to policy, social interaction and environmental changes etc. The main difference appears to be whether the changes in decision making is via changing of internal attributes and parameter values, as described by Perez and colleagues (2002); or whether there are multiple strategies available from which the agent chooses such as in the Lake model by Janssen and Carpenter (1999). This appears to be strongly linked to the type of micro- level decision making as described above.

Another key issue for ABM is that of validation. The logic of simulation as a method is somehow different to that of for instance statistical modelling (Gilbert and Troitzsch 1999). For validation, both methods compare outputs of the method with observed data in order to evaluate similarity (Gilbert and Troitzsch 1999). However, due to non-linearity and unclear causality in the underlying systems, an infinite number of models are able to reproduce a given set of outputs (the equifinality principle). Consequently, similarity between model output and empirical observations is not sufficient for evaluating whether a model is an appropriate representation of reality or not. This equifinality problem can be avoided as described by Gershenson (2002) via logical arguments to dismiss theories that are incompatible with human knowledge. The problem with this is that the logical tests are based on axioms that can not be tested themselves.

By applying social validation, such as in the Companion Modelling approach, stakeholders can convince themselves that the logic and patterns represented in the model are acceptable approximations of the real system. Social model validation is also important for the developing of a sense of ownership and acceptance among stakeholders. On a final note about Agent Based

86 Models, it seems that the technology is now moving towards a mature stage as indicated by moves towards a community framework for Agent Based Modelling (Janssen et al. 2008) that would incorporate what is needed to make the ABM community more efficient and aligned with other disciplines:

• A unified protocol for describing ABMs,

• An online model archive,

• A shared library of model components,

• Shared test beds for models creating the ability for improved validation and testing,

• Better GIS-ABM integration and

• More formal and standardised training and education in ABM.