Scott Moss

Sociology and simulation: statistical and qualitative cross-validation

Agent‐based simulation modeling enables the construction of formal models that simultaneously can be microvalidated against accounts of individual behavior and macrovalidated against aggregate data that show the characteristics of many socially derived time series. These characteristics (leptokurtosis and clustered volatility) have two important consequences: first, they also appear in suitably structured agent‐based models where, like real social actors, agents are socially embedded and metastable; second, their presence precludes the use of many standard statistical techniques like the chi‐square test. These characteristics in time‐series data indicate that a suitable agent‐based model rather than a standard statistical model will be appropriate. This is illustrated with an agent‐based model of mutual social influence on domestic water demand. The consequences for many frequently used statistical techniques are discussed.

From KISS to KIDS: an 'anti-simplistic' modelling approach

A new approach is suggested under the slogan “Keep it Descriptive Stupid” (KIDS) that encapsulates a trend in increasingly descriptive agent-based social simulation. The KIDS approach entails one starts with the simulation model that relates to the target phenomena in the most straight-forward way possible, taking into account the widest possible range of evidence, including anecdotal accounts and expert opinion. Simplification is only applied if and when the model and evidence justify this. This contrasts sharply with the KISS approach where one starts with the simplest possible model and only moves to a more complex one if forced to. An example multi-agent simulation of domestic water demand and social influence is described.

SDML: A Multi-Agent Language for Organizational Modelling

A programming language which is optimized for modelling multi-agent interaction within articulated social structures such as organizations is described with several examples of its functionality. The language is SDML, a strictly declarative modelling language which has object-oriented features and corresponds to a fragment of strongly grounded autoepistemic logic. The virtues of SDML include the ease of building complex models and the facility for representing agents flexibly as models of cognition as well as modularity and code reusability. Two representations of cognitive agents within organizational structures are reported and a Soar-to-SDML compiler is described. One of the agent representations is a declarative implementation of a Soar agent taken from the Radar-Soar model of Ye and Carley (1995). The Ye-Carley results are replicated but the declarative SDML implementation is shown to be much less computationally expensive than the more procedural Soar implementation. As a result, it appears that SDML supports more elaborate representations of agent cognition together with more detailed articulation of organizational structure than we have seen in computational organization theory. Moreover, by representing Soar-cognitive agents declaratively within SDML, that implementation of the Ye-Carley specification is necessarily consistent and sound with respect to the formal logic to which SDML corresponds.

Sociology and Social Theory in Agent Based Social Simulation: A Symposium

A lengthy and intensive debate about the role of sociology in agent based social simulation dominated the email list simsoc@jiscmail.ac.uk during the autumn of 2000. The debate turned on the importance of models being devised to capture the properties of whole social systems and whether those properties should determine agent behaviour or, conversely, whether the properties of social systems should emerge from the behaviour and interaction of the agents and, if so, how that emergence should be represented. The positions of four of the main protagonists concerned specifically with the modelling issues are reprised and extended in this symposium.

Agent-based integrated assessment modelling: the example of climate change

Current approaches to deal with the socio-economic implications of climate change rely heavily on economic models that compare costs and benefits of different measures. We show that the theoretical foundations underpinning current approaches to economic modelling of climate change are inappropriate for the type of questions that are being asked. We argue therefore that another tradition of modelling, social simulation, is more appropriate in dealing with the complex environmental problems we face today.

Understanding climate policy using participatory agent-based social simulation

Integrated assessment models (IAMs) have been widely applied to questions of climate change policy--such as the effects of abating greenhouse gas emissions, balancing impacts, adaptation and mitigation costs, understanding processes of adaptation, and evaluating the potential for technological solutions. In almost all cases, the social dimensions of climate policy are poorly represented. Econometric models look for efficient optimal solutions. Decision making perspectives might reflect broadscale cultural theory, but not the diversity of cognitive models in practice. Technological change is often ignored or exogenous, and without understanding of stakeholder strategies for innovation and diffusion. Policy measures are proposed from idealised perspectives, with little understanding of the constraints of individual decision makers. We suggest a set of criteria for IAMs that can be used to evaluate the choice and structure of models with respect to their suitability for understanding key climate change debates. The criteria are discussed for three classes of models-- optimising econometric models, dynamic simulation models and a proposed agent-based strategy. A prototype agent-based IAM is reported to demonstrate the usefulness and power of the agent based approach and to indicate concretely how that approach meets the criteria for good IAMs and to complex social issues more generally.

Modelling Bounded Rationality Using Evolutionary Techniques

A technique for the credible modelling of economic agents with bounded rationality based on the evolutionary techniques is described. The genetic programming paradigm is most suited due to its meaningful and flexible genome. The fact we are aiming to model agents with real characteristics implies a different approach from those evolutionary algorithms designed to efficiently solve specific problems. Some of these are that we use very small populations, it is based on different operators and uses a breeding selection mechanism. It is precisely some of the “pathological” features of this algorithm that capture the target behaviour. Some possibilities for integration of deductive logic-based approaches and the GP paradigm are suggested. An example application of an agent seeking to maximise its utility by modelling its own utility function is briefly described.

Boundedly versus Procedurally Rational Expectations

Some economists who have relied on the rational expectations hypothesis are now seeking to demonstrate that rational expectations equilibria can emerge in models with agents who are artificially intelligent. They typically model agents’ intelligence through the use of genetic algorithms. However, these algorithms misrepresent current understanding of human cognition as well as well-known and long-standing evidence from business history and the history of technology. This paper implements a well-validated representation of human cognition in SDML, a logic-based programming language that is optimised for representations of interactions among agents. Within that software environment, a model of a transition economy is developed with three production sectors and a household sector. The numerical outputs from that model are broadly in accord with the statistical evidence from the Russian economy. The model itself is developed explicitly to incorporate qualitatively specified characteristics of entrepreneurial behaviour in that economy. Unlike conventional economic models, transactions are negotiated and effected explicitly — there are no unspecified or under-specified “markets”.

Control metaphors in the modelling of economic learning and decision-making behaviour

Learning is represented in economic models either as a process of estimating equations which are known to be correct representations of the environment or as a process of sampling from a known probability distribution. It is arguably more natural to represent leaning as the specification of equations or other relationships in conditions where information processing capacity is not sufficiently powerful as to enable agents to identify such global characteristics of the whole information set as the moments of a probability distribution. Techniques drawn from the literature on machine learning and on knowledge-based systems are coming into use as tools for modelling learning in such conditions. In this paper I describe the key differences between the older and newer approaches to the modelling of learning and decision-making in terms of two metaphors: the menu and the agenda. Two agenda-type algorithms—a genetic algorithm and a production system — are applied to a simple economic model to show that they imply quite different descriptions of learning and decision-making. Moreover, the production system finds the optimal behaviour orders of magnitude faster than the genetic algorithm because — it is suggested — it is the better descriptor of actual strategic decision-making behaviour in normally complex economic conditions.

Canonical Tasks, Environments and Models for Social Simulation

The purpose of this paper is to propose and describe an alternative to an overarching theory for social simulation research. The approach is an analogy of the canonical matrix. Canonical matrices are matrices of a standard form and there are transformations that can be performed on other matrices to show that they can be made into canonical matrices. All matrices which, by means of allowable operations, can be transformed into a canonical matrix have the properties of the canonical matrix. This conception of canonicity is applied to three models in the computational organization theory literature. The models are mapped into their respective canonical forms. The canonical forms are shown to be transitively subsumptive (i.e., one of them is “nested” within a second which itself is “nested” within the third. The consequences of these subsumption relations are investigated by means of simulation experiments.