Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Overview

The previous chapter focused on demonstrating our developed modelling frameworks in the context of one particular case study: Schelling’s well-known residential segregation model (Schelling 1978). Despite this model’s inherent simplicity, the results were seen as significant within social science. Our analysis of the methodological and theoretical underpinnings of Schelling’s model provided some insight into how his simple model of societal micromotives became so influential.

However, Schelling need not be the only example of a successful computational modelling endeavour. While Schelling does fare well when viewed in the context of a number of different modelling frameworks, there are other examples of computational research which can provide useful results through varying means.

By revisiting our central bird-migration example, and viewing each of our developed modelling frameworks from the first two sections of this text in the light of our analysis of Schelling, we can put together a more comprehensive set of ideas regarding the limitations of computational modelling. The effect of such ideas on substantive modelling works are also important to discuss; with these methodological frameworks in place, research utilising computational modelling will necessarily need to adapt to these restrictions.

2 Lessons from Alife: Backstory and Empiricism

2.1 Backstory in Alife

Our analysis of Alife in Chap. 3 focused first on the distinction between two proposed varieties of artificiality: Artificial1, a man-made example of something natural; and Artificial2, something made to resemble something else. This distinction proved most important in the contrast between strong Alife and weak Alife: strong Alife seeks to create systems that are Artificial1 in nature, while weak Alife seeks only Artificial2 systems.

Along with the drive to create digital examples of life, members of the Alife community have sought to use their simulations as a means for providing new empirical data points useful for studying life. Such an empirical goal is far from trivial, and requires a cohesive theoretical backstory to provide a basis for allowing that data to be used as such. As described in Sect. 3.6, our PSS Hypothesis for Life provides one example of such a perspective: if one accepts that life is an information ecology, then a suitably-programmed computer can also instantiate such a system.

However, a backstory of this nature requires additional philosophical baggage. While a researcher may utilise this PSS Hypothesis to justify investigations into a digital living system, that system still exists only in a virtual form. The researcher becomes a digital ecologist, studying the output of the simulation for its own sake.

2.2 Schelling’s Avoidance of the Issue

Our analysis of Schelling provides a means for escaping these philosophical conundrums. Rather than proposing a model which is an Artificial1 instantiation of human society in order to explore residential segregation, he produces a simple model which only illustrates an example of his concept of micromotives and their effect upon society (Schelling 1978).

A version of our bird-migration model could take advantage of simplicity in a similar way. If we constructed a model which presented each bird as only a simple entity in a grid-world, with simple rules which drives the movements of those birds, we may be able to present an example of how singular micromotives could drive birds to shift from one location to another (say by moving according to food distribution or other factors that may contribute to our bird agents’ well-being under the given rule set). In such a formulation the question of Artificial1 versus Artificial2 is unimportant; the model is obviously Artificial2 in nature, and no claim is made to be creating real, digital bird equivalents. The model simply seeks to show the impact of Schelling-esque micromotives in bird migration.

However, while avoiding the problem of artificiality can be advantageous to the modeller, the question of relating that model to empirical data still remains. Without claiming that the agents within our bird model are instantiations of digital life, we cannot be said to collect empirical data on those agents. Might we remain in danger of constructing an artificial world which we proceed to study as a separate entity from the real systems of interest?

3 The Lure of Artificial Worlds

3.1 Levinsian Modelling

Our examination of the more pragmatic concerns of modellers through the lens of population biology provided some additional considerations for the simulation researcher. Levins’ (1966) developed a framework in which three dimensions of generality, realism and precision, important to any given model of a natural system, must be balanced effectively to produce a useful model. He posits that a modeller can focus on two of these modelling dimensions at a time, but only at the expense of the third; this leads him to describe three possible varieties of models, which we denoted L1, L2 and L3 (see Table 4.1).

As noted in Chap. 4, however, applying this framework to certain computational models can be difficult. If our bird migration model uses a highly-simplified set of evolving agents to represent birds, and places those agents in a simplified world with abstracted environmental elements to affect those agents, how might we characterise that model under Levins’ framework? Certainly precision cannot apply, as there is no attempt in this formulation to match these agents with any particular real-world species. Nor does realism seem to apply, as the highly-abstracted nature of the model divorces it from the natural systems it seeks to model. Can the model be referred to as simply general in character, or even then is the model seeking insights which may end up generalising in a different fashion than in other varieties of models?

3.2 Levins and Artificial Worlds

This difficulty leads us to the question alluded to in the previous section: do we risk becoming mired in the study of artificial worlds for their own sake? Certainly our proposed simulation above would suffer from a lack of available empirical data to use in the model. For example, by using evolving agents to represent the birds we can model the progression of this abstracted species over time, but empirically-collected data following a species through its evolution is not available to provide guidance for this aspect of the model. The modeller can proclaim a serious advantage in being able to model something which is impossible to study empirically through only a relatively modest investment of programming time and processing time, but likewise, the lack of relevance to biology becomes ever more acute (once again, see Webb 2009 for a discussion of this issue in relation to Beer’s CTRNN model).

In a certain way, however, this lack of empirical relevance is an attractive feature for the simulation researcher. Replacing the complexities of the real world with simplified abstractions in a wholly artificial world not only makes model construction potentially easier, but even avoids the practical difficulties often associated with traditional empirical methods. As a consequence, our bird model need not be tied down by Levins’ pragmatic modelling concerns, and balancing his three dimensions of model-building suddenly appears much easier.

Of course, as discussed in Chap. 4, such artificial worlds create further methodological difficulties of their own. While such a model may appear to avoid Levins’ concerns and thus produce more complete models of phenomena, confining that model to an artificial world creates a strong separation from the empirical data that can inform other varieties of models. The modeller is thus in danger of studying the model as a separate entity from the systems on which it is based.

Schelling avoids this difficulty by positioning his model as a means to illustrate the importance of micromotives in social behaviour (Schelling 1971, 1978). While his model does relate to a real system, and it does take place in a highly idealised artificial world, within this context the model does not need a strong relationship to empirical data. Schelling instead strives for transparency maximising tractability by creating an abstract, easily computable model. The question of relating the model to the real system thus becomes simplified: can the model illustrate the potential for individual micromotives to exert a great influence on a society? The answer, for most social scientists, appears to be yes, despite the model’s artificiality and simplicity.

4 Modelling in Alife: Thoughts and Conclusions

4.1 Lessons from Schelling

As seen in the analysis above, there are a multitude of considerations relevant to the ALife modeller. From crafting a suitable theoretical backstory to avoiding the difficulties of artificial worlds, methodological problems are hard to avoid completely. Schelling provides some insight into how to approach these issues. The simplicity of the model allows for a coherent theoretical backstory, focusing only on the possible effects of micro-motives on the larger system. Meanwhile, the model’s transparency maintains tractability, though this brings with it a high level of artificiality in the model.

As we see with Schelling, however, this artificiality is not necessarily the problem. The larger concern is the intent with which the model is constructed. A strong ALife practitioner who seeks to create digital life needs to demonstrate Artificial1, and in this case he would presumably require a much higher degree of complexity than in Schelling’s model; even with the PSS Hypothesis for Life as a backstory, a grid-world of simplistic homogeneous agents could hardly be said to compose an information ecology.

4.2 The Power of Simplicity

The circumstance of being driven away from simplicity toward complexity in this search for Artificial1 also creates a much more difficult pragmatic situation for the modeller. As noted in Chap. 7, simple models like Schelling permit the modeller to spend far more time theorising than tweaking. For our bird researcher, a situation in which a few simple lines of encode provide the agent’s behaviour is preferable to one in which each agent contains complex neural network elements, for example. In the first case, the researcher can write the code and run the simulation quickly, then take the necessary time to examine results, run alternative versions of the simulation, and so forth. In the second case, the researcher could spend far more time tweaking the individual parameters of the agent structures: what sort of sensors might each agent have? How do the neural networks control movement and receive input from the environment? How many neurons are there, and what type are they? The questions become many and varied for the researcher using complex agents, and each day spent tweaking those agent parameters is one less day spent pondering the results of the simulation.

As seen above, Schelling’s model provides one example of a means to avoid these pragmatic issues. His model is of such simplicity that writing a version of his simulation takes only a few lines of code compared to more complex simulations. However, clearly other types of models can maintain similar simplicity; Beer, for example, touts his CTRNN-based agents as displaying ‘minimally-cognitive behaviour’ (Beer 2003a,b). Of course, Beer’s analysis of that minimally-cognitive behaviour is extremely detailed and time-consuming, and may indicate that such analysis is impractical for agents of that type even of such relative simplicity. Nonetheless, Beer does demonstrate that relatively simple and analysable neural models are not outside the realm of possibility.

4.3 The Scope of Models

With all of these points in mind, we see that Schelling-type models are hardly the only permissible variety under these frameworks; in fact, a large number of models may display appropriate theoretical back-stories while remaining tractable. Schelling does, however, illuminate the central concerns tied to these modelling frameworks: the importance of theoretical backstory, artificiality, tractability and simplicity, and the scope of the model.

The final element, scope, is an important one to note. Schelling succeeds not only by having a cogent backstory, using its artificiality appropriately, and remaining tractable, but also by limiting its approach: the model aims only to illustrate the importance of micro-motives in a social system, not produce empirical data. In the same way, if our bird migration researcher chose to model the movements of the birds from place to place simply for the purpose of realistic mimicry of their travels, as in Reynolds’ flocking boids (Reynolds 1987), he could do so with little theoretical baggage. If he then chose to declare these mimicked movements as instances of real flocking, or as producing relevant empirical data, suddenly far more justification is required.

5 The Difficulties of Social Simulation

5.1 Social Simulation and the Backstory

While our earlier analysis of ALife provided some valuable insight into the limitations of agent-based modelling techniques, particularly in contrast to traditional mathematical models, these refined frameworks developed in those chapters do not translate simply to social simulation approaches. Within a field such as social science, the philosophical considerations we addressed in those frameworks grow even more troublesome. Imagine that we have created our bird migration model, and the in-depth construction of our programmed agents allows those agents to begin to communicate and even form social structures of a sort. If our stated goal is to use this model to investigate properties of human societies by providing a new, digital source of empirical data, we must not only accept the PSS Hypothesis (presuming that only living things may create a true society), but also related points. The researcher must be prepared to accept that these digital beings, alive but unrelated in a conventional sense to natural life, can create a real societal structure.

In addition, this societal structure will be further removed from real, natural societies by its dependence on a form of ‘life-as-it-could-be.’ In that case, even if we accept that this virtual community of birds can create a society of their own through their interactions, how might we be able to relate that behaviour to the development of real-world societies? If we accept Sawyer’s concerns regarding the non-reductive individualist character of human society (Sawyer 2002, 2003, 2004), are we not placing ourselves even further from a possible social explanation by basing theories upon this digital instantiation of society? Perhaps the non-reductive characteristics of our digital society differ completely from those displayed in human society. Once again, we would be stuck studying the simulation for its own sake.

5.2 Social Simulation and Theory-Dependence

As discussed in Chap. 3, the field of ALife may be considered fundamentally theory-dependent: the structure and function of a given simulation is directly connected to the theoretical assumptions made to create that model. The framework discussed in that chapter chose to deal with this issue by noting the inherent theory-dependence of empirical science, and presenting ideas regarding the importance of theoretical back-stories to any empirically-motivated endeavour.

With social simulation, however, an additional layer is added to this theory-dependent aspect of modelling. Not only do the agents and environment of the simulation present problems of theory-dependence, but also the additional aspects of social communication and behaviour that allows the model to address issues relevant to society.

Further, these additional layers of complexity are not easily subdivided into a hierarchy or other framework which may ease the construction of a computational model (Klüver et al. 2003). The interdependencies between individual action and societal effects means that abstract models will lose a great deal of detail in these missing elements, and conversely that highly-detailed models which address these complexities will veer toward intractability. In either case, theory-dependence becomes a greater issue: abstract models will require strong assumptions to remove these non-hierarchical complexities; and complex models will require incorporating ideas regarding the interaction of individuals and society.

5.3 Social Simulation and Explanation

Even if one does manage to construct a tractable social simulation with reasonable theoretical justification, as noted by Sawyer (2004), social explanation via social simulation is a difficult prospect. While the potential for agent-based models to demonstrate the emergence of social complexities is a possible benefit to social science, whether or not those models can provide a coherent explanation of social behaviour is unclear.

Sawyer argues that societies, like the human mind, display qualities which are irreducible to simple interactions of individuals within that society; despite our knowledge of neuroscience, the higher-level study of mental phenomena (i.e., psychology) is still required to understand the human mind. Similarly, Sawyer argues that human society displays a non-reductive individualism, in which social explanation cannot be complete without addressing the irreducible effects of higher-level social structures. The variation of the bird migration example given in Sect. 5.9.2 gives one example of this phenomenon.

6 Schelling’s Approach

6.1 Schelling’s Methodological and Theoretical Stance

Schelling’s modelling approach, as in our analysis of ALife, provides some important insights into addressing the difficulties of social simulation. Once again, Schelling avoids some of the difficulties facing other social simulations by virtue of the residential segregation model’s simplicity. As a C1 model in Cederman’s framework (see Table 5.1 for a summary), Schelling avoids the methodological difficulties of using complex agent structures (as in C2), or seeking profound emergent behaviours (C3). Likewise, problems of theory-dependence are minimised by using only simple agents with simple rules, with no attempt to address larger social structure. As a consequence, Schelling’s methodology remains influential today with researchers who seek simple models of social phenomena (Pancs and Vriend 2007; Benito 2007).

Schelling further strengthens this approach through his own theoretical framework regarding the best use of models. He posits that general models of behaviour can produce general insights, and that within a given discipline some such insights may not be evident from empirically-collected data. This view seems vindicated by the surprising result of his segregation model, the impact of which was immediate and lasting within social science.

6.2 Difficulties in Social Explanation

However, while Schelling’s model does address a number of concerns relevant to social simulation, the problem of social explanation still presents a difficulty. As noted above, his model takes no interest in the presence or influence of higher levels of social structure; he is concerned only with the actions of individuals. Within the study of residential segregation, his result is likely to be only a partial explanation simply for that reason: housing reform, tax incentives, and other measures designed to influence racial integration in the housing sector are likely to have an impact as well. Schelling of course does not strive for such a complete explanation, as discussed in Chap. 7; however, those who do seek an explanation of residential segregation could try to use Schelling’s model as a starting point.

Unfortunately, Sawyer’s perspective argues that avoiding these higher-level elements in an agent-based model and hoping for them to emerge may be fruitless as well (Sawyer 2004). Even if one were to add additional capabilities and complexities to Schelling’s model, in the hope of allowing for more complex emergent behaviours, an explanation derived from such a model would lack the contributions of higher-level, irreducibly-complex social institutions. As in the birdsong example in Sect. 5.9.2, there is an argument that such elements must be incorporated to produce a complete picture of the development social behaviours. The issue of how to progress from Schelling’s initially successful modelling framework to deeper social explanation thus remains a complex one.

At least the social scientist can take solace in the fact that such concerns are not alien to other modelling disciplines, as in Bedau’s discussion of ‘weak emergence’ as a means for avoiding the difficulty of the effect of downward causation in natural systems from higher-order elements (Bedau 1997). Unfortunately, if anything such difficulties are more acute in social simulation than in biologically orientated simulation, as the additional elements of social institutions, mass communication, and other distinctly societal factors add additional layers of unknowns into an already difficult theoretical situation.

6.3 Schelling and Theory-Dependence

As noted above, the issue of theory-dependence looms large within social simulation, and even for Schelling’s simple and abstract model these problems seem to remain. Schelling’s model is constructed as a singular, vitally important assumption: if we presume that individuals choose their preferred housing based on the racial makeup of their neighbourhood, then segregation will result.

Schelling, and likely other theorists, would argue that such an approach is commendable: Schelling was using his model to test a hypothesis, which is an acceptable role for models. By introducing a highly tractable, transparent model to illustrate the potential import of these factors in residential segregation, he was able to present a new perspective on potential individual choices that can lead to such undesirable social outcomes. However, such an approach becomes difficult when the goal is not hypothesis-testing, but the development of new social theory.

7 Luhmannian Modelling

7.1 Luhmann and Theory-Dependence

Luhmann’s influential treatises on the development of social order (Luhmann 1995) provide a perspective which illuminates potential methods to use social simulation to develop social theory. His ideas regarding the low-level basis for human communication, and the subsequent development of social order, seems a natural partner for the agent-based modelling methodology.

This approach demonstrates the fundamental limitation of Schelling’s methodology alluded to in the previous section. While the residential segregation model can, and does, provide a useful test of a hypothesis which demonstrates the importance of individual behaviour in human society, the overall import of that factor is unaddressed by such a model, given that it is confined to a singular behavioural assumption related to a singular social problem. We may build a Schelling-type bird migration model which demonstrates the importance of certain individual-based factors in driving migration behaviour, but the deeper question remains open: how do these behaviours arise in the first instance?

For the social scientist, these are questions that must be addressed in order to develop a deeper understanding of the origin and development of human society. While we can imagine innumerable scenarios in which a Schelling-type model may illuminate certain singular aspects of social issues and problems, the simplicity and related theory-dependence of the approach limits our ability to investigate the fundamental beginnings of society and communicative behaviour.

7.2 Luhmann and New Social Theory

Luhmannian modelling can address this larger goal of social simulation, clearly a larger goal of the research community (Cederman 2001; Axelrod 1997; Epstein 1999), by removing these elements of theory-dependence. Any model constructed based upon the functioning of human society will incorporate a fundamental theoretical bias, and remove the possibility of investigating the factors that lead our society developing in that way initially.

Doran’s perspective regarding the undefined nature of computational agents in social science (Doran 2000) ties in closely with our developed Luhmannian view. Doran argues that social simulations should begin below the level of agent, allowing for the emergence of agent structures that interact without pre-existing theoretical biases affecting those interactions. In both cases, the removal of theory-dependence from the model is paramount to its success in providing insight for the social scientist into the origin of society.

7.3 Luhmann and Artificial Worlds

Recalling the earlier discussion of Levins (1966, 1968) and the lure of artificial worlds for the modeller, this Luhmannian approach seems to fall squarely within this realm. After all, a simulated environment, not based upon empirical data, which includes abstracted virtual agents is already quite separated from the traditional modes of empirical study. A simulation in which most elements of existing theory and data are removed to study the fundamentals of society seems even further away from the real world; one could easily imagine such a model producing quite unusual agents with idiosyncratic interactions. Once again we return to the prospect of studying a model for its own sake, removed from the natural world.

Where the Luhmannian approach is unique in this regard is the way in which this separation from the real world is vital to the model. The search for a fundamental social theory requires the removal of pre-existing theory-dependent elements in order to produce models which illuminate the importance of pre-societal communications and interactions. In essence, the Luhmannian approach utilises an artificial world to illuminate factors in the real world that we may miss, simply by virtue of our pre-existing biases derived from forming our models and theories within a society.

8 Future Directions for Social Simulation

8.1 Luhmannian Modelling as a Way Forward

Clearly the Luhmannian modelling approach displays promise for those in search of new social theory. Without the removal of more theoretical bias from future social simulations, issues of theory-dependence will continue to provide difficulty for the social scientist who hopes to use these methodologies. Likewise, this approach offers a wide scope for developing new ideas regarding the origin of society, something not provided by Schelling-type methods.

Perhaps more importantly, this perspective brings us closer to the larger goals expressed by proponents of social simulation: a means for understanding the detailed workings of human society. From Levins’ enthusiasm for L3 models (Levins 1966) to Cederman’s for C3 models (Cederman 2001), researchers in various fields continue to seek to explain higher-level behaviour through the interactions of component parts following simple rules. For the social scientist, Luhmann reduces social interaction to its simplest components, giving the community access to a new and potentially stimulating view of the earliest beginnings of human society.

8.2 What Luhmann is Missing: Non-reductive Individualism

Despite these promising elements of the Luhmannian approach outlined thus far, the problem of social explanation remains. Sawyer’s perspective of non-reductive individualism in social science (Sawyer 2002, 2003, 2004) implies that even an elegant portrayal of the origins of society may not be able to allow for the emergence of complex social structures. Given that some of these structures are irreducible to the actions of component individuals in the society in question, there may be great uncertainty as to whether a given set of Luhmannian rules for a simulation may be able to produce such complexity.

Perhaps, then, our comparison with the study of the human mind is more apt that initially thought. As Sawyer notes, there is both a study of mind and a study of neurons (Sawyer 2004); likewise, such an approach is an option for social simulation. Analogous to the study of neurons, Luhmannian approaches can probe the low-level interactions of individuals in a society through the use of models. Then, analogous to the study of mind and mental phenomena, other models may probe the influence of higher-level organisation upon those low-level agents. A combination of these approaches could provide insight into social science that, as Schelling describes, may be unattainable by other means (Schelling 1978).

For example, Luhmann’s discussion of the function of social order (Luhmann 1995) includes an in-depth discussion of the major elements of the modern social institutions which pervade most human societies. This in turn inspired a model in which each of those institutions was modelled very simply, as a monolithic influence on a society of agents, with each institution affecting one element of overall agent behaviour (Fleischmann 2005). Thus, this model attempts to capture the ‘science of mind’ level of societal interaction, while also incorporating lower-level agent behaviours. Such a model shows great promise, as Sawyer’s objections and the difficulties of strong emergence hold much less weight if these downward causation effects can be harnessed appropriately in a model.

8.3 Integrating Lessons from ALife Modelling

As our analysis of Schelling shows, the frameworks in the earlier chapters that underwrite certain varieties of ALife models can be brought to bear on social simulations as well. The issue of theory-dependence has proved vital enough to the success of social simulation that the latter chapters of this text aimed to develop a modelling framework which removes that difficulty. Similarly, the pragmatic concerns of generality, realism, precision and tractability will remain important even in a Luhmannian approach which aims to develop social theory; our modified Levinsian framework provides a useful guide to the limiting factors present in all models of natural phenomena.

A larger question in both ALife and social simulation is illustrated neatly by Schelling’s model and related theoretical justifications: how does the intent and scope of a model affect its usability, either empirically or theoretically? Schelling shows how a simple model, intended to demonstrate the importance of a singular factor in a singular problem, can have wide-ranging effects on related theory. Despite using agents that followed only a single rule, Schelling’s demonstration of the potential impact of individual micro-motives on the segregation problem lead to a great deal of research investigating the impact of such individual factors on all varieties of social problems.

Likewise, our analysis of ALife demonstrated the importance of theoretical backstory in driving the acceptance of a model as a contribution to empirical data within a discipline. Empirical science is full of such back-stories, but they are implicit: the trans-cranial magnetic stimulation researcher believes tacitly that such methods are analogous to producing real lesions in a patient’s brain. In contrast, the agent-based modeller deals with artificially-generated data that is not simply produced in a natural source through a different means, but is produced entirely artificially, in a constructed world. For this reason, the theoretical backstory for agent-based models must be explicit, as otherwise a connection between the model and related data from the natural world becomes difficult to establish.

8.4 Using the Power of Simplicity

As discussed throughout all three parts of this text, the issue of complexity in models has numerous ramifications. For the researcher, highly complex models are difficult to develop successfully; there are a number of choices to be made at the implementation level. For example, an evolutionary model requires a number of parameters to govern the evolution and reproduction of its agents. How should the agents replicate? Should crossover and sexual reproduction be used? What sort of mutation rates might be necessary, and will test runs of the simulation illuminate the best rate of mutation with which to produce interesting results? Additional complexities such as neural network elements or detailed agent physiologies require even more numerous questions at the implementation level.

Similarly, as discussed in relation to the Levins framework, greater complexity leads to greater difficulties in tractability (in the context used here, a greater difficulty in producing substantive analysis, as well as computability). Levins discussed the possibility of producing models of such complexity that they exceed the cognitive limitations of the scientist studying them (Levins 1968). For the mathematical modeller, a model which captures every possible factor in a migrating bird population could end up consisting of hundreds of linked partial differential equations, a mathematical morass of such complexity that analysis becomes fruitless. For the computational modeller, a model of similar character could produce highly-detailed agents with individual physiologies and complex neural structures that allow for remarkably rich behaviour; yet, such a scenario is a nightmare for the analyst, for whom divining the function and impact of these complex internal structures takes incredible amounts of time even for the simplest neural networks (see Chap. 4 for discussion in relation to Beer’s model).

Thus, we must take inspiration from Schelling in this respect. Greater ease in implementation and analysis are two enormous advantages of simpler models. Particularly in the case of social theory, where the potential complexities when studying human society in simulated form are vast, simple and elegant models which illuminate crucial aspects of social systems are the most likely to produce substantive insights.

9 Conclusions and Future Directions

9.1 Forming a Framework

This text has sought to investigate in-depth both the theoretical and methodological concerns facing researchers who utilise agent-based modelling techniques. By examining these issues, and developing frameworks to understand the limitations of such models, future directions for substantive modelling research can be identified.

Our early analysis of ALife in the beginning of the first section of this text demonstrated the importance of theoretical justification in the model-building process, as well as the potential theoretical pitfalls of artificiality in simulations. Whether it is Newell and Simon’s PSS Hypothesis, or Silverman and Bullock’s PSS Hypothesis for Life, computational models require a theoretical basis before they can fit into the conventional tapestry of empirical methods. In the case of both of these frameworks, the implied empirical acceptability of models built on such premises comes at the price of philosophical baggage for the modeller to bear. Those who choose not to take such strong philosophical positions, as in weak ALife, find themselves in the situation of facing more difficult theoretical problems despite avoiding philosophical conundrums.

This baggage became increasingly evident during our examination of Levins’ modelling framework (Levins 1966). The modeller who seeks to produce useful insights about natural systems must strike a difficult balance between both Levins’ proposed three modelling dimensions and the additional aspect of tractability. This fundamentally limits the ability of a researcher to develop models that capture the complexities of natural systems; instead, they are left waiting for techniques that may enhance tractability while still allowing a rough balance of the three Levinsian dimensions.

Initially, our analysis of social simulation in the second section of this text seemed even more problematic than for the ALife community. Social simulations suffer the same methodological and theoretical complexities of ALife simulations, but with the added problem of increased layers of complexity through the addition of social considerations. Even assuming that such problems could be surmounted, the difficulty of providing complete social explanation remained.

However, the introduction of Luhmannian systems sociology introduced another means of utilising simulation in social science. In the case of developing new social theory, creating models that are closer to reality in fact poses a fundamental problem: the issue of theory-dependence prevents the social scientist from producing simulations that avoid biases drawn from our own societal experience. Artificiality thus becomes a desirable trait, bringing the social theorist away from pre-existing theoretical biases in his models. The lure of artificial worlds in this case is thus practical: rather than attempting to avoid the inherent difficulties of modelling illuminated by Levins, and instead losing explanatory capacity, the Luhmannian modeller avoids theory-dependence and gains the ability to generate new elements of social theory.

In essence, the root of the issue comes back to artificiality and intent. A modeller who wishes to create an Artificial1 instance of life or mind can simply look to the PSS Hypotheses. One who wishes to mimic a natural behaviour as accurately as possible, without claiming any theoretical revelations as a result, can simply declare his work Artificial2. A weak ALife researcher, or a similar perspective from other disciplines, can apply his Artificial2 simulation to the examination of a natural system or an element of human society by balancing his model’s dimensions within the modified Levins framework, and ensuring both tractability and a reasonable theoretical backstory (as in Schelling and his focus on general insight and transparency in modelling).

For those who wish to apply their models to the development of new social theory, the issue of artificiality becomes more nuanced. The modeller in this case does not seek an Artificial1 instantiation of a natural phenomenon, nor does he seek an Artificial2 imitation of something natural. Instead, the modeller seeks to develop a simulation in which pre-existing elements of the natural system are removed, along with pre-existing theoretical positions on that system, in order to develop theories independent of bias.

9.2 Messages for the Modeller

Now that we have examined a number of modelling frameworks in detail, analysed and compared each in Chap. 6, and applied them to Schelling in Chap. 7, we have been able to ascertain the most fruitful directions for future models in the social sciences to take. However, on a more general level, the models proposed here will not be the entirety of social simulation. Indeed, a great variety of different types of models will continue to develop in this field, and while Luhmannian modelling may hold great promise for those interested in formulating new social theory and in using simulation by playing to its greatest strengths, this is not to say that other types of models in social science must take a lesser role.

For the social science modeller interested in producing different varieties of models, this argument has brought forth a number of philosophical and methodological frameworks which can provide insight into making those models more relevant to empirical social science. Our discussion of modelling frameworks in Alife illuminated the importance of creating a theoretical backstory; since modelling is inherently a theory-dependent approach, an understanding and elucidation of the theoretical underpinnings of a given model is vital for creating a basis upon which to understand that model. Merely stating that the model is somehow reminiscent of a real-world incidence of some social phenomenon is not enough to create a useful justification.

Our subsequent discussion of modelling in social science saw these frameworks applied to a new area of enquiry, and described the various theoretical problems at issue in the field which have a significant impact for the model-builder. While the idea of a balancing act inherent in creating a model between complexity and analysability is nothing new, bringing the Levins framework into the discussion in social science helps to illuminate more specific pragmatic modelling concerns that are important in computational models. The in-depth discussion of problems of explanation in social science demonstrates how the complexities of modelling human society create additional difficulties for the modeller, adding to the problems of strong emergence in the field of Alife. The modeller interested in social science must take care to understand the limitations of the computational approach, and to avoid incorporating too many elements which may lead to unanalysable complexity. Models offer some new research directions to the social scientist, but also new methodological difficulties; models cannot solve all of these problems and must be deployed appropriately in conjunction with empirical data and useful theoretical background.

Finally, we applied the frameworks discussed in relation to both Alife and social science simulation to one particular example, that being Schelling’s residential segregation model. Schelling provides the best means for demonstrating the tensions between model implementation and model design within social science. The model’s inherent simplicity limits the conclusions which one can draw from its results; yet, that same complexity allows for greater communicability and impact among the social science community. Replicating and extending Schelling’s model is simple in comparison to many computational models, and as a consequence a greater number of the social science community could join in the conversation spurred by the creation and dissemination of the model. In this respect, Schelling’s limitation of his model’s scope to a simple demonstration of the possible impact of a single individual factor on a complex social problem was a big part of its success; the modeller would do well to remember this point as a reinforcement that models need not encompass every element of a given problem to create intriguing insight and new directions for empiricists and theorists.

9.3 The Tension Between Theory-Dependence and Theoretical Backstory

A central element running throughout the three major sections of this text is the discussion of theory-dependence in models. Early on in the Alife discussion this element became important, and the possible theoretical biases on display in a given model become a serious potential difficulty in applying that model to the field in question. The proposed solution to this issue has been the creation of a salient theoretical backstory, one which describes the reasons which make a given model acceptable as contributing in some important way to the discourse of the field to which it is applied. By doing so, the researcher can justify their model and its conclusions as important and not simply relevant only to those interested in mathematical curiosities; as discussed in Chap. 3, the Alife researcher might do so by contending that their model is an instance of an information ecology and thus an instance of empirical data-collection despite its inherently artificial nature.

However, such theoretical backstory can also limit the applicability of a model. Too strong of a theoretical framework, or one which does not hold under closer scrutiny, can result in a model which suffers from greater theory-dependence, rather than less. The tension here then is the delicate balance between providing a theoretical context for a model, and having that model stuck too deeply in theoretical elements that limit its usefulness to the broader field. There is no easy solution to this tension, as these pages have demonstrated. However, the modelling frameworks discussed herein provide guidelines to avoid such pitfalls. A careful examination of the theoretical and pragmatic circumstances surrounding a model can help the modeller to avoid both theoretical pitfalls that may make his model’s applicability suspect (as in Webb’s discussion of Beer Webb 2009), and to avoid those elements of model construction and implementation that may lead to great difficulties in analysis and the presentation of useful results.

9.4 Putting Agent-Based Modelling for the Social Sciences into Practice

Having investigated the use of agent-based models in Alife and social science, we are more aware of the issues facing modellers in these areas and have developed frameworks to help us navigate these complexities. As modellers seeking the most effective application of agent-based models to the social sciences, our work is not yet done, however. General principles for effective modelling are of great value, but within each discipline and sub-discipline of the social sciences there are further complexities to face before we can put these ideas into practice.

Creating individualised modelling frameworks for all the major disciplines of the social sciences is obviously beyond the scope of this text. In order to situate our modelling ideas properly in an individual discipline requires in-depth knowledge of its history, scope and research aims. We would need to analyse the methods of each, determining how agent-based models can enhance researchers’ efforts in that field. Finally, in order to be truly convincing to those looking to branch out into agent-based approaches, we would need to present some worked examples that illustrate the potential of these methods to contribute useful knowledge.

In Part III, we will build upon the foundations presented thus far and offer an example of the development of a model-based approach to a social science discipline – in this case demography, the study of populations. Demography is a field with a long history, and close ties to empirical population data. We will present a framework for model-based demography, a considered approach to demographic simulation that takes into account the needs of the discipline and its practitioners.

In order to develop a model-based demography, we will start by analysing the methodological development of the field, from its earliest beginnings to the bleeding edge. Demography, given its nature and close ties to real-world data and policy, tends toward the social simulation approach to social science modelling. As a consequence, our analysis will focus particularly on the demographic approach to data and how simulation can maintain the field’s focus on real-world populations. We will show that demography not only can live up to demographers’ expectations in this respect, but it can even enhance their ability to gain new insight about the processes and functions underlying demographic change by helping us to address demography’s core epistemological challenges of uncertainty, aggregation and complexity.

We will then examine some worked examples of agent-based modelling, with a particular focus on two modelling projects which will be presented in detail. These two models illustrate how agent-based models can work in tandem with traditional statistical demography to build simulations that closely mirror the behaviour of the real-world populations under study. Finally, we will discuss the status of model-based demography and its potential to contribute to the discipline, and theorise about the implications of this effort at methodological advancement for other areas of social science that are experimenting with a model-based approach.