Order this book
Laboratory of Engineering for Complex Systems (LISC), Cemagref, Clermont-Ferrand, France.
The idea of this book comes from the observation that many works about modelling within the social sciences do not use existing specification languages (like DEVS and UML) to describe their models. The argument for using such a common language is to understand the models better and to understand their structure in the way that it is currently done in other fields like robotics and ecology. The debatable reason the author gives for the current state of affairs in social science is that there is actually no correct reference model to express every model of human behaviour. Professor Schmidt proposes the PECS agent architecture to fill the gap. He argues that the (widely used) BDI architecture can't be considered as a good reference model because it uses rational decision-making in agents as an assumption. Every new framework brings something interesting to the community in the sense that it is another point of view about social systems. Professor Schmidt's book falls into this category of original work that specifies a new framework for modelling human behaviour. This framework is called PECS which stands for Physical Conditions, Emotional State, Cognitive Capabilities and Social Status. The present work can thus be seen as an handbook for this reference model.
Chapter 1 presents a brief introduction to the different uses of "agent" technology which correctly situates the book in the existing literature. The author summarises the uses of agents into three categories (Empirical Science, Engineering Science and Theory) and relates each category to general results and requirements. He situates his own efforts in the Empirical Science field, aiming to produce reference models that require structural similarity with real systems. He then proposes PECS as a reference model for the modelling of social systems.
In Chapter 2, the author discusses general issues about modelling, such as the relation with gathered data, model design and model validation. References are unfortunately lacking but we can find well-known issues about the bias of modellers with respect to the systems they observe. This recovers the definition of modelling as an interaction between two dynamical adaptive systems given in Gell-Mann (1994). The author's position about the definition of a "model", presented only as an abstraction of the real system, fails by not taking into account that modelling is a contextual activity. Minsky (1968) notes on the same subject: "To an observer B, an object A* is a model of an object A to the extent that B can use A* to answer questions that interest him about A". However, building a reference model such as PECS requires abstracting from the context to obtain a generic meta-model. An interesting part of the book is devoted to defining a good modelling process that may be applied more often within the field. The modelling process defined describes the conceptual model design, involving the identification of the model components and the description of the model structure; the model specification (using the PECS reference model); its implementation and its validation by comparing model outputs to data gathered from the observed system. Even if modelling steps like model verification and calibration are not taken into account, it is still a great improvement to apply the techniques of software engineering to modelling activity. Concluding this chapter, the author describes the use of both emergence and reduction in modelling. When modelling individual properties and behaviours, emergence is defined as the phenomenon that leads from individual properties to group formation and dynamics. Reduction is described as the inverse process, linking group properties to individual behaviours and can be used as a valuable means of validation. As the author observes: "If reduction can be done, this is a significant indication that the model is correct".
In Chapter 3, the author justifies his claim that human behaviour can be modelled. Unfortunately, the beginning of the chapter is a little bit confusing. There appears to be a conflation between the fact that one can model the understanding of a human being on a machine and the machine's ability to understand. Another conflation arises in discussing the agent architecture used (PECS) - which may be able to serve as support to every model of human beings - and the underlying wish to create a model of human beings in the first place. Hopefully the paragraph explaining the differences between a model and a replica clearly and humbly defines what a model is and what its limitations are. A replica is defined as "...an identical copy of an original..." and a model as "...an abbreviated depiction of an excerpt of reality based on abstraction and idealisation... It does not claim to be a replica."
In Chapter 4, the author summarises the behavioural models he takes as references. Each of these can be seen either as whole models or as modules that can be co-ordinated to represent a more complex human behaviour. Retaining a classical classification from multi-agent systems, Schmidt sorts behaviour into two major types: reactive and deliberative. The reactive behaviours are categorised into four types. Firstly, instinctive behaviour that follows a simple physical stimulus, state, reaction pattern (reflexes). Secondly, learned behaviour that stands for instinctive behaviour within a social context ("drive on the right"). Thirdly, drive-controlled behaviour that is a reactive behaviour triggered by a physical need (looking for food). Finally, emotionally controlled behaviour that is a reactive behaviour triggered by an emotional state. The deliberative behaviours are defined by goals that the organism tries to achieve by elaborating action plans. They are themselves sub-categorised into two types: constructive behaviour for which the goal is known but cannot be changed and reflective behaviour for which the organism has the possibility of self-regulation. The organism can then set its own goals or it can modify existing goals. A paragraph is dedicated to the discussion of motives corresponding to drives, emotions or acts of will that are associated with precedent behaviours. The level of motives for each behaviour enables to sort the set of possible actions into a dynamical subsumption architecture (Brooks and Connell 1986) that creates the model dynamics.
Chapters 5 and 6 briefly describe the PECS reference model. A more complete description is provided by Urban (2000). As already mentioned, PECS stands for Physical Conditions, Emotional State, Cognitive Capabilities and Social Status. These are the four main building blocks of a particular PECS agent architecture. Adding a Sensor-Perception module and a Behaviour-Actor module, the structure of this agent is described by the author as able to express any model of human behaviour and also as able to replace the Belief-Desire-Intention architecture. It is not possible to enter into a debate on this claim here but it seems to me that even if you are not obliged to define each of these four building blocks for a PECS agent, it is a very massive and static framework. The examples given in the book seem to confirm this intuition. The two rather simple models presented are a little bit difficult to describe using the PECS formalism.
Chapters 7 and 8 give two examples of simple models using the PECS specification to illustrate how one might build models with PECS. The first one, the Adam model is very similar to Sugarscape (Epstein and Axtell 1996) and enables the reader to familiarise themselves with physical, emotional and cognitive PECS components. Within the Adam model, a single agent (Adam) exists on a grid environment with food cells and danger cells. Adam builds a cognitive map of its environment to remember food cells to eat and danger cells to avoid. The physical component contains only energy management (energy grows when Adam eats food, decreases a little when he walks or a lot when he faces a danger). The emotional component manages Adam's fear when it faces danger. The cognitive component manages the building of the map. The behaviour is chosen while taking into account the motives and states of the agent: food needs, fear and so on.
The other example is the Learning Group Model. It illustrates the use of the social status component of the PECS reference model. Within this model, agents are students that have choice to learn alone or together in a number of groups. If they learn in a group, they can choose which group to learn with and apply for it. Agents are defined both by their intelligence (ability to learn) and by their social make-up (the ability to join a group and need for social contacts). Their goal is to acquire the most knowledge they can by learning. The balance between their motives (social satisfaction and learning) and the groups' admission policies - based on such considerations as whether the new member bring learning abilities to the group - drives the dynamics of the model.
I would say that both examples explain the PECS specification and how it can be used to build models in a satisfying way.
The final chapter is very frustrating because it purports to describe a "real" model built with PECS. This model seems so interesting that it is very disappointing not to have anything more than an outline. The model described briefly is the Role-Play Model. It aims to model the role-play of children - Mummies and Daddies, Doctors and Nurses, Cowboys and Indians and so on - and the different group formation, interaction and disbanding processes. However, the author only presents features of the real system, some ideas about the modelling process and about the structure of the model (individual interactions) and states that it mixes the Adam model and the Learning Group Model. This failure to deliver on earlier claims is very disappointing. However in this last chapter, the author also presents an interesting experiment, perhaps inspired by "Big Brother", which is to place children in a room to play and record them with movie cameras. This enables researchers to examine the behaviours under study in detail afterwards.
The overwhelming impression about this book is that it would have benefited from more rigorous completion. The initial chapters, a rather general description of agent technologies, modelling process and modelling of human beings, surely suffer from a lack of references. Existing works and alternative ways of thinking are disregarded, and there is thus no comparison to sustain the arguments presented. In the introduction, the author claims that PECS may replace theBDI architecture but there is a clear lack of argument for this claim. However the book constitutes a good handbook for the PECS specification. It may prove helpful to anyone who wants to review diverse specifications for conceptual modelling, but will probably have no more theoretical impact.
BROOKS R. and J. H. Connell 1986. Asynchronous distributed control system for a mobile robot. In Proceedings of SPIE's Cambridge Symposium on Optical and Optoelectronic Engineering, Cambridge, MA.
EPSTEIN J. and R. Axtell 1996. Growing Artificial Societies: Social Science From the Bottom Up. Brookings Institution Press, Washington, DC and The M. I. T. Press, Cambridge, MA.
FERBER J. 1999. Multi-Agent System: An Introduction to Distributed Artificial Intelligence. Addison Wesley Longman, Harlow. [JASSS review.]
GELL-MANN M. 1994. The Quark and the Jaguar: Adventures in the Simple and the Complex. Little Brown, London.
MINSKY M. L. (ed.) 1968. Matter, minds and models. In M. L. Minsky, editor, Semantic Information Processing, The M. I. T. Press, Cambridge, MA.
URBAN C. 2000. PECS: A Reference model for the simulation of multi-agent systems. In R. Suleiman, K. G. Troitzsch and N. Gilbert, editors, Tools and Techniques for Social Science Simulation. Physica-Verlag, Heidelberg.
Return to Contents of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 2001