[Next] [Previous] [Top] [Contents]

Towards a Descriptive Model of Agent Strategy Search - Bruce Edmonds

2. Computational, agent-based models

The move to agent-based models in economics can be seen as part of a transition to a more descriptive style of modelling. An agent-based model must, by its very nature, model a real actor with a computational agent (in some way), so there should be a one-one correspondence between actors and agents. It is not necessary to assume that the law of large numbers will iron out the messy details. The model can allow the global properties to emerge (or not) without having to assume these details away. Real economic actors are (almost always) encapsulated, i.e. they will have an inside where the decision making is done which is largely hidden from view, and a series of ways in which they interact with the outside environment which are observable. The agents that are used to model these actors are encapsulated in a directly analogous way*1.

However, many agent-based modellers do not see the need for any greater descriptive accuracy than this. Thus when inspecting the learning, inference and decision making processes that an agent uses in such a model, one often finds something as unrealistic as a simulated annealing algorithm or standard genetic algorithm. These are algorithms that have been taken from the field of computer science, regardless of their descriptive appropriateness for the actual economic actors being modelled. Now it is possible that in some circumstances such algorithms will give acceptable results for the purposes of some models, but at the moment we can only guess whether this is the case. It is not only that we do not know the exact conditions of application of each algorithm, we do not know of even a single real circumstance where we could completely rely on any of these `off-the-shelf' algorithms to give a reasonable fit.

To be clear, I am not criticising looking to computer science for ideas, structures and frameworks that might be used in modelling. Being a bounteous source of possible types of process is one of the field's great contributions to knowledge. What I am criticising is the use of such algorithms without either any justification of their appropriateness or modification to make them appropriate.

Thus many agent-based models fail to escape the problems of more traditional models. They attempt to use some ensemble of interacting agents to reproduce some global outcome without knowing if the behaviour of the individual agents is at all realistic. The wish for the `magic' short-cut is still there.

Clearly what is needed is some way of modelling the behaviour of economic actors by computational agents in a credible way. As noted above, real economic actors are probably complex in the sense that it is unlikely that we will be able to deduce their actions from a priori principles. What we can do is to constrain our models as much as possible from what we know. There are several sources of such knowledge.

This is a good start, but when one is using a more expressive formal system like a computational one then this is unlikely to sufficiently constrain the possible models. In other words, there are likely to be many computational model which produce the same global outcomes.

Axtell and Epstein set out some criteria for the performance of multi-agent simulations in [1] In this: level 0 is when a model caricatures reality at the global level through the use of simple graphical devices (e.g. animations or pictures); level 1 is when the model in is qualitative agreement at the global level with empirical macro-structures; level 2 is when the model produces qualitative agreement at the agent level with empirical micro-structures; and level 3 is when the model exhibits quantitative agreement at the agent level with empirical microstructures. The constraint I am suggesting corresponds to their `level 3' with an emphasis on the agreement over time.

Now the task of the cognitive scientist is difficult, but such scientists are able, at least, to exclude some mechanisms for explaining behaviour and make suggestions for the mechanisms derived from a lot of observation. It is notable that many successful sciences take their ultimate grounding for the behaviour of their components from outside their discipline - chemistry is grounded in physics and biology is grounded in chemistry.

This method has its known drawbacks, but can be successfully used, especially when confirmed by other methods. In any case it is likely to produce more useful and accurate information about the real behaviour of actors than is implicit in many commonly used assumptions.

Towards a Descriptive Model of Agent Strategy Search - Bruce Edmonds - 06 SEP 99

[Next] [Previous] [Top] [Contents]

Generated with CERN WebMaker