©Copyright JASSS

JASSS logo ----

Oswaldo Terán (2004)

Understanding MABS and Social Simulation: Switching Between Languages in a Hierarchy of Levels

Journal of Artificial Societies and Social Simulation vol. 7, no. 4

To cite articles published in the Journal of Artificial Societies and Social Simulation, reference the above information and include paragraph numbers if necessary

Received: 28-May-2032    Accepted: 29-Jun-2004    Published: 31-Oct-2004

* Abstract

This paper suggests procedures for decreasing misunderstanding between modellers in social simulation, aiming at helping modellers comprehending a certain phenomena from different perspectives, being aware of the relativity of each approach, and drawing conclusions from the different perspectives. A hierarchy of four levels of language, namely, cultural or natural language, modelling and theoretical paradigm, modelling language, and simulation programming language, is proposed and exemplified as a framework for examining simulation models - assumptions of language embedded in the model at each level are made explicit. Afterwards, switching between languages is suggested for achieving different interpretations and alternative explanations of a model; alongside, as a synthesis from different interpretations, to draw in an interpretive conclusion is suggested. In addition, Interpretive Systemology, a soft systems approach, is proposed as another innovative alternative for better understanding social simulation models, as it recommends undertaking the whole modelling process from different perspectives. The hierarchy of languages, and switching between languages, will be placed against the whole modelling process as understood by Edmonds (2000).

Methodology, Modelling, Social Simulation, MABS, Theory, Philosophy

* Introduction

Several papers have been aimed at explaining better basic modelling notions around the work of the social simulation community. From some of them (see, for instance, Edmonds 2000, Gershenson 2002), it is obvious the lack of transparency in many issues about the social simulation enterprise. It is clear that important matters need to be better explained. One example of such lack of clarity is the discussion around concepts such as social embeddedness, constraint, and emergence, appeared on the social simulation discussion list (SimSoc) (see: http://www.jiscmail.ac.uk/lists/simsoc.html).

Gershenson (2002) addresses issues such as the definition of social behaviour. In addition, motivating problems are identified, for instance, the difficulties appearing when a wide notion of social behaviour is attempted. Such difficulties arise when a researcher attempts to judge the relevance of different papers in the social simulation field without been aware of the different perspectives from which social behaviour is frequently defined in different papers. In general, similar situations might arise as far as the modeller is not concerned with the relativity of key modelling concepts and the various perspectives from which social matters are usually addressed in different papers. This fact brings Gershenson to the following dramatic observation: "modellers might be speaking about different things while using the same words".

In fact, there are different levels of coding taking place along the process of social simulation and the limits among this levels are not delineated clearly enough. The simulation interpretation and analysis sometimes have to take place at different levels and using different languages. It is necessary to be aware that there exist different cultures involved in the simulation enterprise.

The analysis of this paper will point out very important difficulties arising because of the lack of an appropriate contextual framework to compare different models. In order to deal with these difficulties a hierarchy of levels of language and switching between languages in the modelling process, for instance, for interpreting simulation outputs, is proposed.

First, in section 2, modelling in Multi-Agent-Based-Simulation (MABS) is examined to illustrate some difficulties related with the need for making more explicit the modelling context in MABS, and, especially, in social and organisational simulation modelling. Afterwards, in section 3, a framework for social simulation modelling is proposed and illustrated against two modelling tendencies in social simulation. In section 4, switching between languages is suggested as an exercise through which modelling situations can be displayed and better comprehended. Then, in section 5, switching the whole modelling process, an idea similar to that used in Interpretive Systemology (Fuenmayor 1991), is proposed as a procedure to make explicit the whole set of accumulated assumptions while elaborating a simulation model, and to explore alternative modelling process guided by different languages perspectives and assumptions. Finally, in section 6, conclusions and comments about further work are given.

* Examining MABS Modelling: Difficulties in Social Simulation

This section intends to make more understandable the need for a framework aimed at making explicit the origin of the diversity of perspectives in social simulation.

Interdisciplinary Character of the Social Simulation Enterprise

As observed by Edmonds (2000), in MABS the simulation model is a representation of an abstraction rather than of the target directly (as it is the case in human representation in general). The abstraction to be modelled comes from: observation and data collection from the target system, bibliographical review (theories) or consult to expert domains; and, is strongly value-laden. The lumped model or model to be simulated comes from the base model (which is not necessarily an explicit model) after some assumptions and simplifications are introduced. This model, given as a set of interrelated subsystems, contains (explicit or implicitly) "the axioms of the model".

In social simulation not only modelling and simulation theories but also organisational and social theories, as well as metaphors brought in from other areas of research (e.g., ideas of Self-Organised Criticality coming from Physics), may take part in the development of a model. For instance, many organisational models in the literature assume aspects of Cyert's Behavioural Theory of the Firm (Cyert and March 1959, 1963).

Edmonds (2001) points out characteristics of social theories relevant for social simulation. He claims that sociological theory is not a specifically applicable model but "rather a richer sub-language of interrelated terms and ideas useful for characterising and understanding social phenomena". He affirms that it lacks enough specificity as to allow modelling directly. Thus, frequently the simulator creatively constructs a simulation bearing in mind what they know - in this process the modeller adds assumptions. Consequently different models might be elaborated without contradicting a certain theory (some readers might consider that a new theory results when new assumptions are added). As Edmonds indicates, it might be said that sociological theory assists modelling by preventing assumptions that are inconsistent with theory whilst, at the same time, not determining the models. This nondeterminism might be too extensive and, as it is suggested by Edmonds, the linguistic framework given by a sociological theory "may not be definite enough to allow a computational model to be built from it". This nondeterminism gives high importance to the interpretation of a social theory. Interpretation is also highly important in social simulation.

Such inappropriateness of social theories for building simulation models facilitates the inclusion of modelling ideas from other areas of knowledge, and induces modellers to make use of empirical observation.

The influence of this interdisciplinary character of social simulation research permeates in all directions in the social simulation enterprise. Validation is not an exception. Different validation procedures can be found in the literature. One of the more traditional ones, proposed in most simulation and modelling theory books, is by using Monte Carlo techniques.

Alternative validation techniques to traditional ones are proposed by Terán (2001) and by Carley and Gasser (1999). There, enveloping simulation trajectories and proving theorems about the theory of a simulation model are suggested. In this case, ideas come from theorem proving techniques and simulation theory (e.g., the idea of proving) while others come from particular needs observed in social simulation (e.g., the need for observing and studying emergent tendencies in social simulation brings in the idea of enveloping simulation outputs).

Moss and Edmonds (2003) propose cross-validation, i.e., validation at two levels:
  1. Qualitative validation at the level of components, i.e., at the level of agents. Validation at this level occurs when rules of behaviour of the computational agents resemble in some way observed behaviour of the target agents. The idea is to validate the component's behaviour and the components' interaction.
  2. Statistical validation of patterns of behaviour of the overall system. In particular, Moss and Edmond's paper suggests checking the emergence of high volatility (for some variables, - a characteristic of self-organised systems).

Moss et al.'s ideas of validation appears to come from three sources:
  1. MABS modelling and the explicit representation of components resembling social agents, especially that version where the decision making aspects of individuals are represented explicitly as suggested by Newel and Simon - from here comes the idea of qualitative validation of agents' behaviour and interaction.
  2. Ideas and theories from Physics such as that of systems exhibiting Self-Organised Criticality. Hence the ideas of looking for aspects of high volatility in the overall system.
  3. Social theories, particularly the structuration theory, and ideas of social embeddedness. Relating their ideas of cross-validation with social theories helps in supporting the suggested validation procedure.
Different validation procedures generally imply different experimentation procedures - obviously the implemented experiments are dependent on the characteristics of the intended validation.

Language and Context are not Explicit Enough

Edmonds (2000) notices the importance of the "language/theoretical framework" in MABS modelling. He describes the modelling process as constituted by the following steps: abstraction, design (of the model from the abstraction), inference, analysis, interpretation and application. Edmonds argues that the abstraction step consists on choosing a language/theoretical framework and then formulating the abstraction within the chosen framework. He notices that sometimes this step is not given appropriately in MABS modelling (idem, p. 18-19): "instead of the domain (i.e., the target system) determining the most appropriate framework, the formulation is adjusted so that the target system can be mapped into the framework even if this means distorting the target system almost beyond recognition". This happens, he adds citing Kuhn's book The Structure of Scientific Revolutions, because "people tend to see problems in terms of their favourite framework" (idem, pp. 18).

On the other hand, Gershenson (2002) notes the difficulties that appear when the different perspectives simulation models are elaborated from are not taken into account. He gives an example of how this problem appears when comparing several social simulation models without considering the diversity of "contexts" different models are built from and the different goals each model may respond to. In this sense, it is not rare that different models represent abstractions of diverse theories or different paradigms - models are built from the point of view of different experimental frames. Hence the interpretation of theories, even of words, in social simulation modelling should be better considered.

Therefore, if more aspects of the language and context were made more explicit in social simulation this would help to decrease polemic and increase understanding among social simulation modellers.

High Complexity of the Social Simulation Task

The nature of the work of the social simulation community is somewhat different from traditional (e.g., event driven) simulation. It appears that this difference has its source in the fact that in the former the target systems are more complex and the goal is explanatory rather than predictive. While the target of traditional simulation is industrial systems, where the particular human cognitive capacities are oversimplified, the target of the social simulation community is social systems, where usually special attention is given to the decision making ability of the interacting subsystems.

Zeigler's modelling and simulation theory (Zeigler 1976), as well as his idea of experimental frame, is intended for studying "hard" rather than "soft" systems. His definition of experimental frame starts describing the simulation context from the objectives of the study. It does not touch other aspects of the modelling enterprise such as: the paradigm or language containing the theoretical framework; the theory itself; or the simulation programming language. This is so because assumptions about such aspects within the community of researchers working in event driven simulation are relatively similar. Such issues are not currently under discussion and the polemic about them is very low. Thus, they are less critical and deserve less attention than in the social simulation enterprise, where a more suitable modelling framework appears to be necessary.

Given the high complexity of the social simulation enterprise, descriptions of the modelling context should include, in addition to those aspects considered in Zeigler's definition of experimental frame: (a) a hierarchy of languages such as the language of the theory (the paradigm), the theory itself, and the simulation programming language; as well as, (b) other aspects of the context such as the validation procedure and the experimental design.

An Answer for These Difficulties

The interdisciplinary character of the social simulation enterprise, where several research areas such as social theories, physical theories, expert domains knowledge, data from the modelling system, modelling theories, etc., participate in the elaboration of a model; the ambiguity and interpretation problems; as well as the high complexity of this enterprise, demands a suitable "framework" where the whole modelling situation could be displayed and better comprehended. Such framework would help social simulation researchers to understand better what they do and to place their particular work in a better characterised context. We suggest a hierarchy of levels of languages as that framework (in section 3), and switching between languages as an exercise through which modelling situations can be displayed and better comprehended (in section 4). The hierarchy of languages will be placed against the whole modelling process as understood by Edmonds, highlighting the interpretation task, which is highly important in social simulation and, in general, in social research.

* A Framework for Social Simulation Modelling

To deal with the difficulties pointed out in previous sections, I suggest that the following hierarchy of levels of language should be made explicit (a level may consist of several languages). Each language is supposed to be used for a certain human community in a certain context and with some purpose. A language at a certain level is supposed to be used in a wider and richer though more loose and difficult to characterise "human environment" or context than a language at a lower one.

More specifically, the contexts of each language at the hierarchy are given by: the social interaction where the every day human interrelationships occurs and a culture is displayed, for a language at the higher (first) level; the community of scientists undertaking a specific theoretical and modelling paradigm, for a language at the second level; the community of scientists adopting some particular modelling approach, for a language at the third level; and, the research community assuming the simulation perspective given by a simulation programming language, for a language at the lower level.

It is supposed that any phrase in a certain language can be somewhat interpreted in another language at a higher level in the hierarchy. i.e., a phrase in a simulation or in a modelling language would have a certain interpretation, probably by using common sense, in our every day language. In a logically defined hierarchy of languages a phrase in a language at a certain level could be somewhat "translated" into a phrase in a language in a higher level but the converse would not be necessarily true[1]. In the proposed hierarchy of levels of languages, for languages at different levels, this relation is assumed true to some point but is not the aim of this classification to choose excessively fixed and logically constrained relations among languages.

Languages co-evolve, as the influence between languages in terms of change in one language requiring change or "revision" on the other language, such as adding of new words or meanings, is not necessarily static and downwards. For instance, changes in a modelling language might make, to some point, a paradigm (a language in a higher level) irrelevant and might help in triggering a new and more appropriate paradigm. For instance, observations that an individual's behaviour is context dependant and that its decision-making mechanism "operates" by using problem spaces leads to a modelling language which describes individuals as heterogeneous decision-makers; this, in turn, affects the (paradigm) assumption in traditional Economics that individuals are homogenous and have perfect knowledge about their environment; finally, this contradiction brings in new paradigms in Economics, where heterogeneity among individuals and an individual's bounded rationality are recognised. Such changes might result from the implementation of ideas coming from a scientific area different from that of the simulation and modelling community. This is helped from the interdisciplinary character of social simulation.

A language at a certain level has sense in a context offered by a language in a higher level. For instance, a simulation programming language is understood and recognised as useful in the context given by a (or some) modelling language(s).

A Hierarchy of Languages and Contexts

  1. Natural Language and Culture. This level refers to the every day language in a culture (e.g., the Latin-American, the Anglo-Saxon, the German)[2]. The interest is in the relevant aspects concerning the simulation and modelling tasks. This is taken as the highest level since languages at the lower levels seem to be subsumed by languages at this level - a natural language, less precise but more flexible, is useful for describing a wider range of situations than languages at lower levels.
  2. Theoretical and Modelling Paradigm. This level refers to the general guiding principles of the scientific perspective, in our case to the modelling and simulation perspective (e.g., Simon's notion of individuals' bounded rationality in organisational and economic modelling, or Einstein's General Relativity Theory in Physics). A theoretical and modelling paradigm gives more general guidelines than a modelling language - several modelling languages are plausibly enclosed by a single theoretical and modelling paradigm. It is exactly because of its impression and flexibility that it is useful for a wide range of contexts; thought, it still depends upon abstraction and therefore simplification. Given the interdisciplinary character of social simulation, there is not a single and well-known paradigm. In an interdisciplinary study a paradigm corresponds to a particular perspective taken elements from different research areas. The theoretical and modelling paradigm would enclose aspects of the social theory or theories under which the model is developed or from which the modeller takes elements. In particular, social paradigms differ in their ontology and epistemic about a social system, and are related to different sociological traditions. Modelling paradigms might present these differences too, but they seem more concerned with the epistemic issues - with the how of the modelling task.
  3. Modelling Language. This level of languages refers to languages used for describing the model, under the guidelines of a certain paradigm, bringing in ideas from different sources and scientific disciplines. Usually, in social simulation, the model is related to social or to organisational systems of knowledge, to modellers' experience, to knowledge of expert domain, to simulation and modelling theories, and to observation. Thus, the language at this level would be given in part by social and organisational theories but also by the modellers' experience, by the advice of domain experts, by observation, by a particular simulation and modelling tradition (e.g., by a MAS, event or process oriented simulation), or by a social or organisational theory (e.g., by a behavioural theory of organisations as Cyert et al.'s). This level of languages refers also to more concrete, specific and operative ideas than those in the theoretical and modelling paradigm level (e.g., Cohen and Moss notion of endorsements for implementing, in part, Simon's bounded rationality idea). In this paper, this level of language is called the Modelling Language, though it might also be called the Theoretical Framework/Background/Context.

    It is not a surprise to find simulation models built up from or taking elements from more than one social theory. Typically, a simulation model does not follow a social theory strictly, the important point is that the model does not contradict the fundamental assumptions of a (some) social theory(ies). This is important in order to make the computational model applicable to a wide range of contexts, as usually social theories are narrow and usually contradicted in the real world - what is understandable in case of any approach trying to understand complex systems like society.

    Other aspects of the context of a model such as the goals of the study, the experimental frame (as defined in Zeigler 1976), and the descriptive model, might be identified at this level of languages.

    At this level, the model validation and experimental design are described and justified in terms of: a) theories, observation and expert domain knowledge used to build up the model; b) the modelling and simulation perspective; and, c) the goals of the study.

  4. Simulation Programming Language. In social simulation, different simulation programming languages constraint the simulation model in different ways, as they usually differ in their assumptions. Differences are not only in representation but also in how they make inferences. Examples of general assumptions are those made when a particular inference mechanism is implemented in a language, or more particularly those about the order the simulation rules are "fired" (or carried out). Thus, at this level the axioms of the simulation model are drawn from the axioms of the descriptive model after aggregating assumptions required and allowed by the simulation platform. Some researchers have considered the question of how different (or how close) two simulation models representing a single abstract model but built in different platforms are (see Axtell et al. 1996).

* Using the Hierarchy of Languages to Examine some MABS Models

In this section, several MABS models are examined in order to make clear aspects of their context. The chosen models represent characteristic tendencies in social and organisational modelling. The idea is to make clearer the diversity of approaches and points of view, and to show the usefulness of the hierarchy of levels of language to make explicit their differences. It is expected that differences between two models at a certain level of language predispose towards differences at the lower levels of language. First two important trends in social simulation will be examined, and following, in order to show a more detailed analysis, a specific model will be reviewed. Issues concerned with the highest level of languages (natural language and culture) will not be considered in this discussion, as the "polemic" at this level is low within the social simulation community.

Next will be explained why the level of natural language and culture is included in the framework above and, however, it is not involved in the discussion. Such level is included in the framework because this is attempted to be general, useful not only in social simulation but also in other research areas. On the other hand, it is on involved in the discussion because there is not high disagreement (either implicitly or explicitly) among social simulation scientists at such level of languages. This might be explicated by the fact that the work of the social and organisational simulation community has sense within a scientific culture more or less similar for all researchers, institutionalised and established in a post-modern society. Disagreement in many issues, for instance, about the main social and organisational problems of the present "epoch", might arise if a different point of view from the dominant one (related to post-modernism) were taken. For instance, it might happens that the understanding and suggested solutions for a forest management problem at the Brazilian Amazon given under a cultural perspective of a Brazilian society aiming at an "endogenous development" or of a certain Amazonian tribe would not be straightforwardly comprehensible from the point of view of social researchers embedded in a post-modern society. Therefore, the perception of a problem and its solutions might have sense in a certain culture and tradition but not in another, and thus disagreement at the highest level of languages might arise.

An Overall Overview: Examining Two Social Simulation Trends

The two trends are represented as follows:
Case 1:
Social and organisational models emphasising the agents' decision making mechanism. For instance, the classical models of Cohen et al. (1972) and Masuch and LaPotin (1989), and the more recent models of Carley et al. (1998) and Moss (1998).
Case 2:
Social models involving cellular automata or grid words. For instance, the models of Epstein and Axtell. (1996) and Nagel et al. (2000).
Second level of language: Theoretical and Modelling Paradigm

In social simulation, unlike what happens in already mature scientific disciplines, it is not easy to find well-defined paradigms. In spite of this, modelling trends indicating important aspects of paradigms can be found. Two trends are identified each related to one of the two cases considered above: related to case 1 is that trend represented by Simon's followers, and, linked to case 2 is that tendency given by researchers using cellular automata models.

The main elements of the paradigm in social and organisational modelling as assumed in the models of case 1 is concerned with the seminal work (initiated in the 1940s) of researchers that includes authors and papers such as Newell (1990), Newell and Simon (1976), Simon (1948, 1984), Cyert and March (1959, 1963), Cohen (1985), March and Simon (1958). The approach centres its attention on the decision-making process and behaviour of the individuals and the overall system (e.g., of an organisation). Important postulates are: individuals are rationality bounded, both the individual's surrounding and the organization's environment are complex (e.g., technology is unclear), the individuals' decision-making situations are ill-defined (there is ambiguity of choice and problematic preferences).

An individual's decision-making process is considered a key aspect for understanding social and organisational systems. Cognitive theories of individuals (Newell 1990, Simon 1984, Cohen 1985) have been suggested and implemented. Case 1's models take elements from these developments. For instance, Moss (1998) assumes the idea of endorsements (suggested in Cohen 1985) and the idea of space of problems (suggested in Newell 1990).

Another important aspect of this paradigm is the rejection of theoretical-biased approaches (armchair theorising) as an alternative for social science (Simon 1986, Edmonds 2001). They claim for the need to give observation a preponderant role in the modelling process, in accordance to the state of social theory at present (Edmonds 2001).

Recently researchers working under this paradigm have pointed out the convenience that social simulation models reflect important properties found in empirical systems such as high volatility of some variables and other properties related with self-organised criticality, and social embeddedness (Moss and Edmonds 2003).

On the other hand, the cellular automata models (case 2) give more attention to emergent properties of the system as a result of the agents' interaction without caring too much about the agents' decision-making mechanism. Consequently, aspects such as the haziness of the decision situations are not an important issue. A suggestive difference between the two paradigms is that in case 1 both the rules of behaviour of each agent and the population of agents are allowed to evolve, while in case 2, evolution is mainly concerned with the population of agents.
Third level of language: Modelling Language

Aspects of a model to be considered at this level are the assumed social, organisational, modelling and simulation theories and methodologies, as well as modelling ideas coming from other scientific areas. Also important are the goals and experimental frame, as well as the validation and experimental design.

Obviously, the modelling language shown in case 1's models should take ideas from Simon and followers. In particular the models of Cohen et al. (1972) and Masuch and LaPotin (1989) follow Cyert et al.'s Behavioural Theory of the Firm. In the former model the modelling approach is not too much elaborated, while in the later model ideas from artificial intelligence and symbolic logic are introduced - for instance, agents (as actors) are represented explicitly and the idea of space of problems is brought in. Masuch et al.'s model implements the ideas of bounded rationality, which was not present in Cohen et al.'s model. The two models share an objective: to explain how organisations can survive despite pervasive apparent disorder. The models have similar experimental designs - in both of them Monte Carlo techniques are used.

Similarly Moss' model and Carley et al.'s model take on important aspects from the named Behavioural Theory of the Firm. Moss' model also uses domain experts advise for elaborating the cognitive model of the agents. Experts' advise is used mostly to set up parameters of the cognitive model. Likewise, Carley takes assumptions from the named organisational and cognitive theories.

For the other case in point, the cellular automata models, social theories do not have a high impact in the elaboration of the model, apart from giving ideas related to emergent properties of the target system and to certain general aspects of the agent's interaction. Similarly, cognitive theories do not have an important role in building the model. Nevertheless, in cellular automata models found in the literature ideas from other disciplines (especially from Biology or Physics) have an important impact in the modelling language. For instance, properties of Self-Organised Criticality are searched for in the simulation dynamics (Nagel et al. 2000).

It is common to find models sharing the modelling paradigm but differing at the level of the modelling language or at another lower level of language. Two models might share Simon and partners' paradigm while the aspect of interest might vary among, for instance: cooperation, coordination, adaptation, imitation, or self-organisation. In this situation, aspects of the language at the third level, the modelling language, would differ.

As a more concrete example of this, consider the following two models: (1) Gershenson's (2001), where social agents are modelled from an autonomous agents (or behaviour-based) perspective (e.g.,), and (2) Castelfranchi's (1998), where social agents are modelled from a rational agent perspective. For the two models the ability of the subsystems for symbolic manipulation is highly important. However, there is a "minor" difference between them: in (1) the interest is in modelling social reasoning, while in (2), the interest is in modelling social adaptive behaviour. This small disparity is consequence of assuming different social behaviour notions, what in turn is related to discrepancies with respect to the assumed social theory, the goals of the study and/or the experimental frame. Nevertheless, the two models might well be following the same modelling paradigm.
Fourth level of language: Simulation Programming Language

In case 1, simulation programming languages should permit to implement the agents' structure and their interactions, as well as the agent's cognitive mechanism as suggested by Simon, Newell, Cohen, and others. An example is SDML (Moss et al. 1998) used in Moss (1998), which provides facilities for social simulation.

Simulation programming languages used for cellular automata modelling (case 2) are potentially more theoretical free, as the modelling language usually requires fewer assumptions than in case 1 in relation to theoretical social and organisational constrains. It is common to find cellular automata models built in general purpose languages such as Java.

Examining A Specific Model: Moss' Critical Incident Management Model (Moss 1998)

Elements pointed out, about Moss' model, in the more general analysis above will be examined in more detail in this section. Before reviewing the model in terms of the hierarchy of levels of language, the paper's abstract and key aspects of the adopted modelling perspective will be described.

Moss' paper abstract states:
The main purpose of this paper is to demonstrate an empirical approach to social simulation. The systems and the behaviour of middle-level managers of a real company are modelled. The managers' cognition is represented by problem space architectures drawn from cognitive science and an endorsements mechanism adapted from the literature on conflict resolution in rule based systems. Both aspects of the representation of cognition are based on information provided by domain experts. Qualitative and numerical results accord with the views of domain experts.
Brief description of the model

The distributed structure of the model is assumed by the modeller and validated against descriptions and data provided by a manager. Managers also describe the nature of critical incidents as well as the likelihood of an incident of one kind leading to incidents of other kinds.

A distributed representation of information and knowledge is facilitated by the agent (container) hierarchy of SDML (see Moss 1998), the Multi-Agent System used to build the model. It allows representing knowledge common to everyone in the organisation and knowledge common among specific groupings of individuals (e.g., departments). In addition, SDML presents facilities enabling, and someway delimiting, social interaction.

In the following the main agents forming the structure of the model are shown. Some of the agents function just a modules. In addition the main aspects of the structure of the organisation, particularly those referred to the agents' interaction and information transmission, can be seen in Figure 1., reproduced from Moss' paper.

Figure 1
Figure 1. Agents' interaction in Moss' model

In Moss' model cognition is based on ideas from ACT-R and SOAR, and on managers advice. It aims at modelling managers' learning concerned with diagnosing and remedying problems creating critical incidents. It is represented by using problem space, endorsements and, by mental-model-sharing when an agent is allowed to copy rules from the rulebase of other agents

Endorsements are tokens that have associated numerical values. For instance, the instances of type Controller endorses their own mental models (decision-rules) with any or all of the following tokens:


For instance, a new (formulated) mental model is endorsed as "newModel"; or a mental model reducing the number of critical events in the next time period is endorsed as "reducedEvents". The total endorsement value E is calculated as

Equation 1 (1)

where b is an arbitrary number base not less than 1, and ei are the tokens.

The mnemonic endorsement tokens and the choice of b are determined by the modeller considering domain experts' advice.

Specialisation and generalisation of models is considered. For instance, a specialised mental model of two successful mental models results from taking the union of the set of causes specified by the two models and the union of the predicted effects.

A Controller filters the models it retains actively in memory over time by remembering from one day to the next only those models with weights of 2 or more. It selects a model to apply by: first, devising a list of candidate models and then choosing the best of the candidates; and, second, choosing randomly a candidate model, by assuming probabilities for a mental model proportional to the mental model's total endorsement value. In the next subsections, Moss' model will be reviewed under the three lower levels of language.
Second level of language: Theoretical and Modelling Paradigm

Aspects of the modelling paradigm are listed below (the central assumption is in italics). Some assumptions are more explicit than others in Moss' model, in the sense that they are easier to identify or are explicitly stated. Precisely, one of the benefits from this analysis is to make explicit the (implicit or) less explicit assumptions, decreasing misunderstanding among social simulation modellers. The more explicit assumptions are listed first:

Less explicit assumptions are:
Third level of language: Modelling Language

As said above, a modelling language alludes to more specific and particular guidelines than a theoretical and modelling paradigm. It is important to have in mind that the modelling language allows to build a descriptive model, under the guidelines of the theoretical and modelling paradigm. Here, aspects of Moss' modelling language, enclosed by the constraints imposed from the above described theoretical and modelling paradigm will be described. The following elements are identified (once more central assumption are in italics). The more explicit assumptions are listed first:

Less explicit assumptions are:

This discussion aims at making clearly the differences between the two considered levels of language, i.e., the modelling and theoretical paradigm and the modelling language, and especially reviewing how guidelines[3] or paradigm assumptions imply constraints upon the modelling language.

As an example of how a guideline or a paradigm constraint brings in related constraints at the level of the modelling language, consider the guideline: Environment is Considered Complex. A first related modelling language assumption in Moss' model is that agents are not allowed to know the inter-events cause-effect relationship - as a consequence the decision-making situation is Uncertain. Other related modelling language's suppositions are: Need for Communication, Task Oriented, and Distributional Constraint. These consequences clearly correspond with Simon's partners and followers' modelling tradition.

Likewise, in line with the guideline: Empirical Oriented Modelling, avoiding armchair biased modelling, many aspects (assumptions) of Moss' model at the level of the modelling language come from expert domain (including managers) advice. For instance, assumptions respect to the structure of the model, such as the choice of the relevant agents and their interrelationships, or respect to the inter-events causal relationships.

All characteristic listed until now including those described as part of the paradigm, namely Bounded Rationality, Path Dependence, Empirical Oriented Modelling, Environment and Agents are Considered Complex, Emphasis On the Decision-Making Process, Mental Models Evolve, Constraints About the Mental Models (Use of Symbolic Logic, Space of Problems and Experimental Psychology), Interest on Emergent Properties, Need for Communication, Information Ubiquity, Agents are Task Oriented, Distributional Constraint, Uncertainty and Agent-based Modelling, form part of a perspective in organisational modelling: the Information Processing perspective.

All these identified elements of Moss' modelling language include some elements more related with theoretical issues and others more concerned with modelling aspects. For instance, the Behavioural Theory of the Firm includes many of these elements. This theory also includes elements not cited here. Nevertheless, such elements do not contradict Moss' assumptions. One example is the conjecture that, in an organisation, usually an individual's preferences are unclear, ill defined, and inconsistent.

In relation to the modelling issues, named assumptions are related to a computational modelling language evolving since Cohen and March's (1972) Garbage Can Model. This model has been fed (and constrained) with ideas coming from the Behavioural Theory of the Firm, and from researchers such as Simon and Newell. Such ideas are somewhat operationalised and implemented in a modelling language. Assumptions in Cohen et al.'s model are "evolved" as Masuch and LaPotin (1989) builds a second model, adding suppositions such as: (a) symbolic logic is a superior device for modelling social systems than imperative programming as it allows, for instance, to represent discontinuous problem spaces; (b) actors or decision-makers should be represented explicitly; (c) and object oriented programming is superior comparing with previous programming paradigms (from 1989). Latter, new assumptions appear, for instance, that Multi-Agent Systems are a good device for representing decision-makers - advancing the idea of object oriented programming and explicit representation of agents. Carley and Gasser (1999) show in more detail aspects of this modelling perspective.

Similarly, elements of a modelling language related to representing cognition have been evolving. Once more, original ideas go back to Newell (e.g., SOAR) and Simon's seminal work (for instance, the paradigm notion of bounded rationality). Much of this work is related to general guidelines or part of the paradigm described in the previous section, such as the idea of representing an agent's cognition by using simple rules even in case of a complex environment; the preference for Experimental Psychology, over other approaches such as Physiology, for investigating about individual's decision making mechanism; or, representing cognition by using production rules. Further work became more connected with operational issues. For instance, Moss' description of managers' cognition by using endorsements.

Moss' model includes assumptions related to the goals of the study, its operational constraints or experimental frame (Zeigler 1976, explains the idea of experimental frame), as well as with the validation and experimental design. Moss model of the North West Water incident management organization is aimed at understanding and somewhat resembling the decision-making process and learning of managers, and the overall organisational response to critical incidents. This goal constrains the modelling task, as it indicates which aspects of the organisational phenomena are relevant for the model. For instance, the particularities of the repair team's work are not considered in detail, while high relevance is given to information transmission among agents and to explicit representation of agents' mental models concerned with decisions about actions to cope with critical incidents. Thus, when experimenting, statistical measures of overall behaviour should be directed at measuring effectiveness of organisational behaviour in terms of reducing critical incidents. This, in turn, indicates the kind of variables that must be explicitly represented in the model. All this is connected with the goals of the study and with the validation procedure. Validation includes observing correspondence between tendencies in the overall organisational system and on the agents' behaviour. However, not only quantitative validation in terms of, for instance, statistical measures, is of interest in this sort of models, but also qualitative validation. Moss investigates how duration of critical incidents changes over time (overall behaviour) and how an agent's mental model of inter-events relationship changes over time. This is concerned with the above named paradigm guideline calling for giving special attention to emergent properties of the system.
Fourth level of language: Simulation Programming Language

In the same manner a modelling language allows to build a descriptive model under the guidelines of a theoretical and modelling paradigm, a simulation programming language allows to elaborate and experiment with a computational model under the guidelines of both, the theoretical and modelling paradigm, and the modelling language. Evidently, all these three languages are subsumed in a cultural language.

Language assumptions in Moss' model include SDML's Underlying Logic, Syntax of Rules, and Internal Manipulation of Constraints; as well as the particular Facilities for Social Simulation. In the following these aspects will be shortly and partially described. Most of theses assumptions are implicit in Moss' paper.

* Switching Between Languages

The previous section shows how a modeller can make clearer the languages/context of a model under a certain perspective. Switching between languages would be a step forward for decreasing the difficulties arisen when a researcher attempts to judge the relevance of and/or compare different papers in the social simulation field, as it allows being aware of the different perspectives from which a certain situation is addressed. It helps a modeller in understanding a certain phenomena from different perspectives, being aware of the relativity of each approach, and drawing conclusions from different perspectives.

Modelling is usually described as having both forward and backward (corrective) steps. For instance, after attempting to design the simulation model it might be decided to modify in some way and for some reason the descriptive model. Here we are not interested in describing the backward counteractive steps, which, however, are assumed to be present.

As said above, Edmonds describes the modelling process as constituted by the following steps: abstraction, design, inference, analysis, interpretation and application. It is possible to design a descriptive model and a simulation model. Both are of interest in social simulation: the first more pure in terms of correspondence to the abstraction and the second more biased and constrained by a simulation programming language. Because of the usefulness of these two descriptions, in this section the model design step is considered to consist in two sub-steps: a descriptive model design step and a simulation model design step.

The whole simulation process, from abstraction to application, vs. the hierarchy of languages, and switching between languages from the interpretation step downwards, is depicted in Figure 2. There, arrows indicate the modelling sequence and the effects between steps in such sequence, especially in terms of aggregating assumptions and reductions as the modelling process advances - for the first two steps sources of such assumptions are explicitly shown. The idea is to make clearer each step, and specially those high in "polemic" in social simulation - such as analysing and interpreting simulation results. As the figure shows, different interpretations might be considered, each in accordance with diverse languages (e.g., several modelling languages), but also a single interpretation drawn in from all these interpretations, a sort of second order interpretation, might arise. This sort of second order interpretation will be called interpretive conclusion. This suggests a novel way for interpreting and analysing simulation outputs.

Figure 2
Figure 2. Switching between languages for interpreting simulation results. Languages might be, e.g., at the Modelling Language level

A more general framework is shown in Figure 3: the whole modelling process against the hierarchy of levels of language is illustrated. In this figure, arrows between steps in the modelling process indicate aggregating of assumptions and reductions, arrows between the levels of language stand for the inclusion relation between levels, and arrows between a level of language and a step in the modelling process specify influence of such level of language upon the modelling step, as the level of language implies assumptions in the modelling step. For instance, the simulation programming language does not imply assumptions upon the abstraction. However, an arrow does not indicate that any constraint from the language level can be added upon the pointed modelling step. For instance, Operations Research has been sometimes criticised because "the modelled situation has been adapted to an applied technique" rather than the other way around. In this case, assumptions from a level of language related to a specific technique (e.g., linear programming) have been imposed over the abstraction, constraining this in such a way that it does not represent the modelled situation appropriately anymore.

Figure 3
Figure 3. Influence of the Hierarchy of Levels of Language on the Whole Simulation Process

It is important to make clearer that a natural language description will likely lead to a range of models based upon a range of interpretations of the real world - interpretation occurs frequently in human understanding in general not just at the point of interpreting model output. Figure 3 shows that usually only the three higher levels of language affects the two first simulation steps, namely abstraction, and descriptive model design. The simulation programming language, along the languages at the other levels, influences downwards from the simulation model design step. At any step of the modelling process alternative results might appear if different languages either at the same or at different levels are used (e.g., different design models at the second modelling step, or different interpretations at the sixth step).

Interpreting is highly important in social simulation and deserves special attention. Switching between languages helps in achieving different interpretations and alternative explanations of a model (moreover, interpretation is possibly the most important stage in any realm of science or human understanding, not just in social simulation). Interesting situations might occur, for instance, it might happen that a certain property of a model is interpreted as emergent under a language but not under another one.

There are many questions related with switching between languages and interpreting that demand answers, these will be subject of further work. They include:

* Interpretive Systemology: Switching the Whole Modelling Process

At each step of the modelling process described above, assumptions are added in accordance to the constraining languages, while a reduction happens. For instance, in the abstraction process, abstraction-assumptions are added (many implicitly) as the modelled phenomenon is simplified in accordance to a "modeller background" (language) and goals. For instance, usually a Sociologist would model a certain phenomenon in a different manner from an Engineer, even the components of the system would be understood and assumed differently. After the whole modelling and simulation process is completed, the accumulated set of assumptions comprises: {abstraction-assumptions, design-assumptions, inference-assumptions, analysis-assumptions, interpretation-assumptions and application-assumptions}. Obviously, for one sub-step (e.g., interpretation) different and alternative subsets of assumptions might be used, as said above.

Interpretive Systemology (Fuenmayor and López-Garay 1991), a soft systems methodology, suggests undertaking the whole modelling process from different perspectives. For our case, it would mean following the modelling process several times, from assuming different abstractions and different abstraction-assumptions on. In such a case, different sets of assumptions would be accumulated for each modelling exercise, each giving a different model. Like switching between languages, Interpretive Systemology advances the idea of using the hierarchy of levels of languages to uncover the modelling assumptions, as this suggests to draw conclusions from different models, each following different assumptions for the whole modelling process. A conclusion in Interpretive Systemology comes then from different interpretations, each interpretation suggested from a different model, models being elaborated from different languages, including the abstraction. On the other hand, an interpretive conclusion, as explained above, is elaborated from different interpretations of the simulation analysis (see Figure 2). Interpretive Systemology already has shown to be a valuable approach for understanding organisational and social systems.

Implementing such a procedure, i.e., switching the whole modelling process (as Interpretive Systemology suggests), helps a modeller in understanding better the modelled situation, uncovering his assumptions in a modelling process, and being aware of other modellers' points of view about the modelled phenomena.

* Conclusions

This paper has offered procedures for helping a modeller in being aware about the relativity of key modelling concepts and the various perspectives from which social matters are usually addressed in different papers and from different lines of research. Such procedures assist in making clear the different levels of coding taking place along the process of social simulation and in delineating more clearly the limits among such levels.

They are aimed at lessening misunderstanding between social simulation modellers. Several causes for such misunderstanding have been analysed: Interdisciplinary Character of the Social Simulation Enterprise, Language and Context are not Explicit Enough, and High Complexity of the Social Simulation Task.

The first procedure suggests a hierarchy of levels of language for analysing simulation models - assumptions of a model would be made explicit at each level of language. The hierarchy ranges from culture and natural language to the simulation programming language, going through the intermediate levels that include the theoretical and modelling paradigm and the modelling language. This hierarchy is used for analysing, first, two modelling trends in social simulation, and, second, a particular model in the first tendency. The analysis shows how the whole modelling situation could be displayed and better comprehended by using such framework and how it helps social simulation researchers to understand better their task and to situate their particular work in a better characterised context. There, not only the modellers' explicit assumptions but also the implicit ones are listed, making clearer the modellers' perspective. A summary of the conclusions from analysing the two modelling trends, namely Case 1: Social and organisational models emphasising the agents' decision making mechanism, and Case 2: Social models involving cellular automata or grid words, is shown in the next table.

Modelling CaseCase 1Case 2

Language Level
Second Language Level:
Theoretical and Modelling Paradigm
- Emphasises individual's decision-making process and behaviour;
- Considers individuals rationally bounded;
- Recognises that the individual's surrounding and the organization's environment are complex;
- Rejects theoretical-biased approaches;
- The rules of behaviour of each agent and the population of agents are allowed to evolve in the model.
- Emphasises emergent properties of the system as a result of the agents' interaction without caring too much about the agents' decision-making mechanism. Aspects such as the haziness of the agent's decision situations are not an important issue.
- Only the population of agents is considered evolving.
Third Language Level:
Modelling Language

Ideas come from:
- Simon partners and followers;
Behavioural theory of organisations;
- Artificial intelligence and symbolic logic;
- Knowledge representation such as the notion of space of problems;
- Implementations of the idea of bounded rationality.
Neither social nor cognitive theories have an important impact in the elaboration of the model. Social theories simply inform about emergent properties of the target system and about general aspects of the agent's interaction. Ideas from other disciplines (especially from Biology and Physics) have an important impact.
Fourth Language Level:
Simulation programming language

Simulation programming languages should permit to implement the agents' structure and social interaction, as well as the agent's cognitive mechanism as suggested by Simon, Newell, Cohen, and others.Simulation programming languages are more theoretical free (e.g., Java), as the modelling language usually requires fewer assumptions in relation to theoretical social and organisational constrains.

The second proposed procedure recommends switching between languages. It implies a step forward in the use of the hierarchy for displaying and better comprehending modelling situations, as each switch will allow considering a different interpretation (and perspective) of the modelled situation. Its exercise permits a modeller to assume other modellers' perspectives, and/or suppose different perspectives by himself. It is also proposed to use this switching for developing novel simulation methodologies, of which one is recommended: considering different interpretations and attaining interpretive conclusions.

The third suggested procedure is Interpretive Systemology, a soft systems methodology, suggesting undertaking the whole modelling process from different perspectives. Alike switching between languages, Interpretive Systemology represents a step forward in using the hierarchy of languages for analysing social simulation work.

* Acknowledgements

The research reported here was funded by the CDCHT (the Council for Scientific, Humanistic and Technological Development) of the Universidad de Los Andes, Venezuela, under project I-524-AA. I would like to thank anonymous JASSS referees and the Editor, Nigel Gilbert, for useful advice in reviewing a previous version of this paper.

* Notes

1 In addition, a level of the hierarchy of levels of language can enclose several languages, thus it is possible to talk about a “quasi-hierarchy” of languages. Many parallel hierarchies of languages/contexts might exist, which can then be described in a single hierarchy of levels of languages.

2 The idea of language here includes not only words and grammar, but also meaning.

3 Here and in the next sections, the word “guideline” means a constraint from the assumed paradigm.

* References

AXTELL R, Axelrod R, Epstein J, and Cohen M (1996) "Aligning Simulation Models: A Case Study and Results". Computational Mathematical Organization Theory, 1(2), pp. 123-141.

CASTELFRANCHI, C (1998) "Modelling Social Action for AI Agents". Artificial Intelligence 103, pp. 157-182.

CARLEY K and Gasser L (1999) "Computational Organization Theory". In WEISS, G. (ed.), Multiagent Systems: a Modern Approach to Distributed Artificial Intelligence.

CARLEY K, Prietula M, and Lin Z (1998) "Design Versus Cognition: The Interaction of Agent Cognition and Organizational Design on Organizational Performance". Journal of Artificial Societies and Social Simulation, 1(3) http://jasss.soc.surrey.ac.uk/1/3/4.html.

COHEN, P (1985) Heuristic Reasoning: An Artificial Intelligence Approach. Pitman Advanced Publishing Program, Boston.

COHEN M, March J, and Olsen J (1972) "A Garbage Can Model of Organisational Choice". Administrative Science Quarterly, 17(1), pp. 1-25.

CYERT R and March J (1963) A Behavioural Theory of the Firm. Englewood Cliffs, Prentice Hall, NJ.

CYERT R and March J (1959) "A Behavioural Theory of Organisational Objectives". In Haire M. (ed.), Modern Organization Theory (pp. 76-90), Wiley New York.

EDMONDS, B. (2001) "Against Pure Theorising". Published at the third section of: CONTE R., Edmonds B, Moss S and Sawyer R (2001) Sociology and Social Theory in Agent Based Social Simulation: A Symposium. Computational and Mathematical Organization Theory, 7(3), pp. 183-205.

EDMONDS, B. (2000) "The Use of Models - Making MABS More Informative". MABS2000 @ ICMAS-2000: The Second Workshop on Multi Agent Based Simulation, Boston, July 9, 2000. Published in: MOSS S and Davidsson P (eds.) (2000) Multi Agent Based Simulation (MABS-2000). Lecture Notes in Artificial Intelligence, Vol. 1979, Springer Verlag, Berlin.

EPSTEIN J and Axtell R (1996) Growing Artificial Societies: Social Science from the Bottom Up. The Brookings Institute Press, Washington D.C. and the MIT Press, Cambridge, Mass, USA.

FUENMAYOR R and López-Garay H (1991) "The Scene for Interpretive Systemology". Systems Practice, 4, 401-408 (see other Fuenmayor's papers in the same issue).

GERSHENSON, C. (2001). Artificial Societies of Intelligent Agents. Unpublished BEng Thesis. Fundación Arturo Rosenblueth, México.

GERSHENSON, C. (2002) "Philosophical Ideas on the Simulation of Social Behaviour". Journal of Artificial Societies and Social Simulation 5(3), http://jasss.soc.surrey.ac.uk/5/3/8.html

MARCH J and Simon H (1958), Organisations. New York, Wiley.

MASUCH M and LaPotin (1989) "Beyond Garbage Cans: An AI Model of Organizational Choice". Administrative Sciences Quarterly, 34, 38-67.

MOSS S and Edmonds B (2003) "Sociology and Simulation: Statistical and Qualitative Cross-Validation". CPM Report No.: 03-105, http://cfpm.org/cpmrep105.html.

MOSS, S. (1998), "Critical Incident Management: An Empirically Derived Computational Model". Journal of Artificial Societies and Social Simulation vol. 1, no. 4, http://jasss.soc.surrey.ac.uk/1/4/1.html

MOSS S, Gaylard H, Wallis S, and Edmonds B (1998) "SDML: A Multi-Agent Language for Organizational Modelling". Computational Mathematical Organization Theory, 4(1), pp. 43-69.

MOSS, S, Edmonds B, and Wallis S (1997) Validation and Verification of Computational Models with Multiple Cognitive Agents. CPM-97-25, http://www.cpm.mmu.ac.uk/cpmrep25.html.

NAGEL K, Shubik M, Paczuski M and Bak P (2000). "Spatial Competition and Price Formation". Working Paper Number 00-05-029, Santa Fe Institute, Santa Fe, NM.

NEWELL, A (1990) Unified Theory of Cognition. Harvard University Press, MA.

NEWELL A and Simon H (1976) "Computer Science as Empirical Enquiry: Symbol and Search". Communications of the ACM, 19(3) 113-126.

SIMON, H (1986) "The Failure of Armchair Economics" [Interview]. Challenge, 29(5), pp. 18-25

SIMON, H (1984) The Sciences of the Artificial. MIT Press, Cambridge, Mass., USA.

SIMON, H (1948) "The Proverbs of Administratio". Public Administration Review, 6(June), pp. 53-57.

TERÁN, O (2001) Emergent Tendencies in Multi-Agent-based Simulations using Constraint-based Methods to Effect Practical Proofs over Finite Subsets of Simulation Outcomes Doctoral Thesis, Manchester Metropolitan University, Manchester, UK, http://papers.ssrn.com/sol3/papers.cfm?abstract_id=292408.

ZEIGLER, B (1976) Theory of Modelling and Simulation. Robert E. Krieger Publishing Company, Malabar, FL, USA.


ButtonReturn to Contents of this issue

© Copyright Journal of Artificial Societies and Social Simulation, [2004]