Reviewed by
Michael
Möhring and Klaus G.
Troitzsch with the collaboration of Markus Bartz,
Michael Brozek, Anh-Phuong Do, Jens Götz,
Jeff Licker, Tobias Nold and Achim Steffes.
Institute for Computer
Science Applications in the Social Sciences, Department of
Computer Science, University of Koblenz-Landau, Germany.
As the editors state in their preface, this volume is based on papers accepted for the Second International Workshop on Multi-Agent-based Simulation (MABS-2000) which was held as part of the 4th International Conference on Multi-Agent Systems (ICMAS-2000) in Boston in July 2000. So at first sight this seems to be just another volume of workshop proceedings. The editorial efforts, however, make it a reader in multi-agent based simulation which can successfully be used in a university seminar to give students insights into multi-agent simulation, into doing their own simulations, into criticising scientific papers, and writing their own. Thus, besides its original purpose, namely to give an overview of current research in MAS, it serves four additional purposes.
This review originates from a seminar held at our university in the winter term of 2001/2002 when seven students had to read Scott Moss's Editorial Introduction and Bruce Edmonds's paper on The Use of Models: Making MABS More Informative and one of seven other papers in the volume. These they had to report on and review, mainly in the light of section 7 of the Edmonds paper that enumerates eight groups of criteria that reports on multi-agent simulation should meet. Thus this review is restricted to the seven papers selected for the seminar and, of course, to the introductory papers by Moss and Edmonds.
The paper by Edmonds starts with an attempt at defining what MABS could be - among others it could be illustration, mathematics, communication and science. We shall come back to these uses of MABS later on. An analysis of modelling, the modelling process and MABS as a modelling paradigm follows, all of these sections are short and precise, and quite appropriate for students preparing for an examination on social science modelling and simulation. Another section gives an overview of some archetypes of MABS papers, illustrating this overview with references to papers in this volume. The most important section of this paper (at least in our eyes) is section 7 where Edmonds lists more than thirty questions (criteria) in eight groups "that can help us judge work, one for each of the modelling steps". He notes, of course, that "it would be impractical for all these steps to be covered in every paper". These criteria are very helpful for authors of papers on social simulation (not only agent-based simulation). The paper ends with a call for a "structured archive of MABS/SS papers" - a task which could be performed by a learned society devoted to encouraging the international advancement of theory and research based upon social simulation.
The Editorial Introduction by Scott Moss puts the MABS paper into the ICMAS context and then classifies the papers submitted and selected for this volume, first into two classes - those exploring "how agent interactions can be used to build multi-agent systems" and those which "offer examples of problem-oriented (rather than technique-oriented) systems" - and then into a larger number of classes - of which we selected for this review from those papers which "draw explicitly and directly on sociological theory", from those "grouped under the applications heading [which] are all examples of problem-driven multi-agent simulation models" and from those which "address issues of actual social interaction in order to develop clear hypotheses for empirical validation".
Both introductory papers not only show the effort which the editors have made to produce a volume of proceedings which is also a very nice reader in the topic, but in addition give a short and concise introduction to the field as a whole.
The papers we selected for our seminar are the following (in the order in which they appear in the volume) - apologies to the authors whose papers we do not discuss in this review:
Thus, the application areas of these papers range from multi particle systems in mechanics and anthropology to sociology and integrated assessment.
The Virtual Food Court paper describes the applicability of a modelling tool (DECAF) to "economic modelling, including the modelling of voluntary organisational contracts". Thus the paper is about a very abstract target system which is then made more concrete by introducing a virtual food court. The VFC serves as an economy in a nutshell which is described in sections 3 and 5 of the paper. Sections 2 and 4 describe the authors' interests - "explaining how organisations come about out of large numbers of autonomous individuals, in a changing environment, with changing technologies and changing demands". These sections also describe their tool, DECAF (Distributed, Environment-Centred Agent Framework). In our view, this paper is at least as much about the applicability of a tool and about tool building as it is about a target system modelled with the help of this tool. This is, of course, a legitimate rationale for writing a paper and for including it in a volume like this; but tool building is not within the scope of the criteria proposed by Bruce Edmonds.
The Simulated Road Traffic paper is about methods "used to solve conflicts that can happen between agents that represent simulated drivers in simulated road traffic". Here we have a cleardescription of a target system; the design of the simulation model is described more or less implicitly. Inferences from simulation runs are compared to real world data. Thus this paper does not only belong to the class of problem-driven simulations (where Scott Moss puts it), but at the same time to the class addressing issues of empirical validation.
The Sand Piles paper is devoted to the "physical problem of granular piling". The target system is a grain silo where under some conditions "the outflow is often stopped by the creation of granular arches" that are difficult to break. The abstraction to a "simplified physical model" is clearly described: the three-dimensional problem is reduced to a two-dimensional one and further restricted to a torus to avoid boundary conditions. The design of the simulation model is documented by both the mathematical formulas describing the simplified model and some pseudo-code examples giving an impression of the implementation.
The Simmel paper goes back to Georg Simmel's On Fashion and to the "'Simmel effect', i.e. persistence of social differences under instability of status symbols, as an effect of imitation and distinctiveness". The abstraction here comes from Simmel's work and is converted into the design of a cellular automaton. This design seems to have been implemented in a quite straightforward manner. A number of experiments is reported and analysed statistically.
In the Sexual Attraction paper, the class of target systems is the set of group-living primates in which "males are dominant, but in spite [of] this dominance [...] allow females access to resources during the period when females are sexually attractive - but only then and not otherwise". Thus the abstraction is quite clear from the very beginning: the paper studies "whether in the absence of benefits and of calculative intentions, female dominance and, consequently, male 'tolerance', increases by self-organisation more strongly when males are attracted to females than when they are not." The second section specifies the design with some implementation details (which would even allow a replication) and describes the experimental set up. Section three reports a variety of results with well-commented diagrams. The final section discusses the results - with "several unexpected consequences".
The Renewable Resources paper deals with "the way interacting agents should understand their environment so that a common good used by the whole group would last". The target systems that the authors have in mind are described very briefly in section 1.3 - they are African communities of hunters, gatherers, nomad shepherds and farmers. Different abstractions follow from these target systems, and different designs follow from the abstractions - which are documented with diagrams, some of which refer to the design, others refer to results. We took a deeper look at the JuMel model which "captures the situation of herdsmen trying to have access to a resource over time" where the resources are owned by local farmers. We did not find a clear distinction of abstraction and design (but this mixture of abstraction and design is presented with all necessary the detail) and no hint of implementation at all. Results - some of them again unexpected - are clearly described: the "friend priority" choice is superior to a "cost priority" schedule. The general discussion at the end of the paper reveals that this effect could also be observed in the other three models: "that competition has a bad impact on the production".
The Climate Policy paper discusses integrated assessment issues and the applicability of agent-based social simulation to these issues. The paper discusses a prototype model originating from the FIRMA project. This model deals with "the consequences of climate change for such water issues such as drought, flood, changing use patterns due to irrigation needs and the like"; this prototype model "is sufficiently coarse grained and abstract as to be applicable to other regions and issues". So we have a clear description of the target systems in mind and of the abstraction. The design of the model is presented, with some hints at the implementation, and a number of results are discussed. Beside the material results, the authors refer also to the methodological issue of "collaboration between the ABSS and the IAM communities [which] holds substantial advantages for both.
The seven examples which we discussed in more detail (though still too little) reveal the apposite nature of the remark that "it would be impractical for all these [modelling] steps to be covered in every paper". For one thing, all the papers in the volume are restricted to twenty printed pages (and some are even shorter) and this is too short to discuss all modelling steps in such detail that it would be possible to replicate the study. In addition, however, all seven papers focus on different parts of the modelling process and come from different backgrounds. To return to the attempted definition of MABS proposed by Edmonds: some use multi-agent based "simulation as a stand-in for symbolic deduction on distributed systems where such deduction is impractical". This is especially true for the Sand Pile paper and perhaps also for the Simulated Road Traffic paper. Some use "multi-agent systems as a tool for understanding observed systems". This seems to apply to the Renewable Resources and Sexual Attraction papers. Some use them "as an interactive medium for social exploration, negotiation and communication" which is especially true for the Climate Policy paper with its focus on stakeholder participation. Finally, some are "designed to animate or otherwise illustrate some sociological [...] principle - see the Simmel and Virtual Food Court papers. (None of the papers in the volume seem to be of the Entertainment and Art types also mentioned by Edmonds.)
At the end of the discussions in our seminar we found that all the papers analysed failed to give answers to all of the questions Edmonds asked in his paper - which did not come as a surprise. On the other hand, all the papers provided all the necessary answers, given the few pages allowed to them. What our students tended to complain about - they are students of computer science - was mainly the fact that the step from design to implementation was only roughly described in all cases. However, this is probably also due to the restricted space in a proceedings volume. The same is true for missing answer to the question about whether "the outcomes are critically dependent upon particular parameter settings". This was another recurring complaint from our students. This is mainly due to the fact that most papers describe work in progress (and we cannot wait for the final results of simulation project, we also want discuss their early stages). But all these complaints should not be taken too seriously. Our students admitted that they had learnt a lot in reading, reporting and reviewing the selected papers: a lot about multi-agent simulation, a lot about reading and reviewing scientific papers, and, last but not least, a lot about writing simulation reports. All other readers will certainly draw the same profit from reading this book. In the end, we can conclude our review with a recommendation to use this book in more university seminars on social simulation. We can also recommend authors describing their simulations to take the Edmonds criteria very seriously - as all of the authors of this volume have done - in order to improve communication about multi-agent simulation.
Return to Contents
of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 2002