Reviewed by
Pietro Terna
Dipartimento di Scienze economiche e finanziarie
Università di Torino
Italy
If you want understand reality you can use agent based simulations. If you want understand an agent based simulation you can play with it or read about it (or, even better, you can use it).
A good example to play with can be found clicking on "The Game" at IcoSystem.
From their site:
"[We can use] The Game to provide a practical demonstration of the following points:
Simulation is a powerful tool for understanding the dynamics of complex systems.
- Simple rules of individual behaviour can lead to surprisingly coherent system level results.
- Small changes in rules or in the way they are applied can have significant impact on the system level results.
- Intuition can be a particularly poor guide to prediction of the behaviour of complex systems above a few levels of complexity (here we have only 3).
You can play The Game for real with 10 or more participants. Ask everyone each to select 2 individuals randomly - person A and person B. Now ask the participants to move so that they always keep A in between themselves and B - so that A is their protector from B. Everyone in the room will mill about in a seemingly random fashion and will soon begin to ask why they are doing this. Now tell them to stop, and that they are now the protector so tell them to move so that they keep themselves in between A and B. The results are striking. Almost instantaneously the whole room will implode on itself with everyone clustering together in a tight knot.
By using a simple agent-based simulation in which each person is modelled as an autonomous agent following the rules, one can actually predict the emergent collective behaviour (see The Game). Also, by using the simulation as test bed, one can explore the design of the rules to produce a desired outcome."
Indeed, with this simple example (try it!) you can directly reach the core of the agent based simulation framework: each agent is represented by an independent piece of computer code; only the interaction among the agents produces the emergent phenomena that we want analyse.
While playing, we also discover that multi-agent simulation can be oriented to reproduce, on the basis of simple and mainly indirect rules, the complex structures of the real world. This is the case for the naturally inspired computing found at http://www.rennard.org/alife/english/antsgb.html, where agents describe and (most interestingly) reproduce the behaviour of ants.
From the site authored by Jean-Philippe Rennard:
"An ant is quite a simple animal. Its behavioural repertory is limited to ten to forty elementary behaviours. Yet, anthills are very complex. One can find nursery, warehouses or kitchen gardens. Some individuals forage, others take care of the eggs, repair the nest or protect the anthill against miscellaneous threats. What is the mystery, how can so mindless animals achieve such a complex organisation?
Division of labour could be the key. Ants are highly specialised, so specialised that some individuals have to be fed by others, they are unable to get food by themselves. In economy, division of labour means efficiency, but to work properly, it supposes some supervision, the different tasks have to be co-ordinated. Yet such a supervisor doesn't exist in anthills, no ants (and particularly not the queen) are able to manage this exploit. Nevertheless, the co-ordination necessarily exists; it results from some self-organisation process.
Let us examine foraging strategies in ants to exemplify this idea.
At the beginning, a number of ants are walking, more or less randomly outside the nest. They are looking for food. All along their way, they deposit a light trail of pheromones. When an ant finds some food, it gets back home, deposing a stronger trail (the intensity of the trail possibly depends on the richness of the found resource). Since ants have trail-following behaviour, a growing number of individuals will tend to follow it and to reach the food. When they get back, they reinforce the trail. A positive feedback (auto-amplification) therefore appears, more individuals reinforce the trail, attracting new individuals which will at their turns reinforce the trail...
In this example, the ants don't communicate directly. Information are exchanged through modifications of the environment (here local gradients of pheromones). [...] The nest structure itself co-ordinates the workers tasks essentially through local pheromone concentrations. The state of the nest structure triggers some behaviours which then modify the nest structure and triggers new behaviours until the construction is over. The process is similar in ants foraging.
The ants tend to follow pheromone trails, but it is only an inclination. There is at any time a positive probability for the ant to abandon the trail and to move more or less randomly. It is then possible that the "lost" ant finds a new resource, eventually far richer than the one that was previously exploited. By constructing a new trail, this ant will attract new individuals and a new positive feedback loop will be set up.
Finally, when satiety occurs or when the resource is empty a negative feedback loop appears. For example, if pheromone decay is quick enough, when the resource is over, less and less ants will tend to follow the trail which will progressively disappear."
The well organised and well planned book which is the object of this review is a good starting point to read about multi-agent systems (or, according to my preferred expression, agent based simulation).
In the first part, Janssen explains that real systems are never in equilibrium and that complex adaptive systems cannot be predicted accurately, in contrast to the traditional Newtonian paradigm, to which economics strictly adheres.
In Chapter 4, the key point is exposed: we cannot predict the future of social-ecological systems and we can only study conditions under which emergent phenomena hold. This is the only informed way, though necessarily humble, we can define our goals, tools and resources in making choices.
With these simple words we can completely summarise research in the field of multi-agent studies.
A wide range of complex adaptive systems are illustrated in this volume, including immune systems and nervous systems as well as economic and ecological systems. Different contributors use complex adaptive systems to give insights into the circumstances in which providing simple local rules can lead to emergent macro-level structures.
The second part of the book is partly devoted to a discussion of the validation of multi-agent systems. The volume tries to link artificial agents to experimental data, whereby the validity of different phenomena are tested. As the authors explain, multi-agent models are widely applied in different ecological settings such as agricultural, forest, and range land systems in Africa, Asia, America, Australia and Europe. The applicability of multi-agent models to the study of ecosystem management is thus clearly illustrated. The usefulness of multi-stakeholder involvement in developing, discussing and applying models to solve contemporary ecosystem management problems (including dealing with their very complexity) is also strongly emphasised.
The quality of ecosystems is affected by the actions of different stakeholders who use them in a variety of ways. In order to understand this complex relationship between humans and nature, it is vital to understand the complexity of the interactions among agents and the authors attempt to do this by applying multi-agent systems. The multi-agent methodology for ecosystem management is a relatively new and rapidly developing field which takes a formal computational perspective on the interaction of humans with their environment.
Chapters 2, 3 and 4 of the book provide the underpinnings from the methodological point of view.
In the second chapter by Anderies, we find the description of a framework for the transition from local interaction to global dynamics via multi-agent models, with a useful formal representation. Anderies also underlines the difference in interaction hypotheses between multi-agent models (with explicit interaction rules) and differential equations type models, where it is implicitly assumed that each agent interacts with every other.
In Chapter 3, Janssen addresses the key problem of rules. In the perspective he presents, rules can be seen as an outcome of a continuous process of tinkering, experimentation, making errors, breaking rules, selection and competition.
Bradbury's Chapter 4 is devoted to futures, predictions and other foolishness: the deep connection between predictions and models (and science) is investigated specifically from the point of view of building and understanding agent based models.
I have summarised in Table 1 below the various contributions to the volume. I have also provided a personal judgement about the difficulty of replicating the simulation results. Obviously, one can always ask the authors to send the code used in their simulations, but this simple request can create a lot of problems. From my own experience I found that when I did not attach the code I used (and the instructions for running it) to my publications, I had enormous problems replicating even my own results of only a few years before. The computer environment does not remain the same, supporting files are sometimes lost, systematic conditions may change after computer recovery from crashes and so on. So, it would have been better if the editor of a such an important book had obtained the code from the authors.
| Author/s | Content | Perspective and results | Difficulties in replicating the results |
| J. M. Anderies | From local to global dynamics | Theoretical framework | Does not apply |
| M. A. Janssen | Self organisation and institutions | Theoretical framework | Does not apply |
| R. Bradbury | Unpredictability of social-ecological systems | Complex system behaviour | Does not apply |
| S. M. Manson | Validation guidelines and tools | Applied methodology | Does not apply |
| W. Jager and M. A. Janssen | Agent based simulation and laboratory experiments interaction and synergic utilisation | Applied methodology | Replication is possible but not easy |
| M. A. Janssen et al. | Agent based improvement of differential equation models | Applied methodology | Replication is possible but not easy |
| A. Balmann et al. | Agent based simulation in agriculture | Applied simulation | The methodology is fully accessible but replication is not facilitated |
| G. Deffuant et al. | Agent based simulation of research information propagation and of the related effect on decision making | Applied simulation | Replication is very difficult |
| T. Lynam | The relationship between household characteristics and long term production effects | Applied simulation | Replication is not easy |
| M. Hoffmann | A multi-agent model of deforestation and reforestation | Applied simulation | Replication is quite easy: C++ code is available from the author on request |
| F. Bousquet et al. | Simulation of collective decision-making processes | Applied simulation | Replication is not easy |
| N. Abel et al. | Collective decision and land use | Applied simulation | Replication is quite possible but not easy |
As always from the perspective of computer code, the book fails to address the Babel problem with regard to different simulation languages. The Tower of Babel can be seen as a symbol of innovation and the search for novelties, but the lack of a lingua franca for the field of agent based systems remains a difficult challenge.
Finally, I return to my key point about the use of multi-agent models to study conditions under which emergent phenomena hold.
There are currently attempts to apply economic concepts (like utility optimisation and price system interpretation) to organise systems in different (non-economic) contexts. An example is the project to develop a "Grid Economics".
From the perspective of the decision-making process in organisations (for example) this approach is very promising because where prices do not operate at all (and classical economics has very poor explanatory capabilities) complexity means complication. Moreover, the other fields of social science do not provide such a strong framework to understand, explain and modify organisational activities.
But ... with a handbook like the one edited by Janssen, it is necessary and urgent to underline that these new approaches need to use economic ideas within a simulation based on a complexity framework and not alone. Direct mathematical solutions to complex situations simply do not exist either in the real world or in artificial ecologies like the Grid.
... Let the play begin!
Return to Contents
of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 2005