Reviewed by
Wojtek Przepiorka
Utrecth University
Standard economic theory, to remain analytically traceable, makes simplifying assumptions about economic agents that are obviously wrong. Making simplifying assumption is in line with any type of modelling technique, be it in chemistry, architecture or psychology; models help us to understand the world by reducing it to the essentials. But the necessity to remain analytically traceable has been contested with the steady increase in computing power. The last three decades have seen the rise of computational approaches to biology, medicine, sociology, etc. Although computer simulations have been used in these disciplines since the invention of microprocessors, the spread of microcomputers with graphical user interfaces and object-oriented programming languages have paved the way to its popularisation in the form of agent-based approaches (Chapter 2 presents a review of this development). “If you can’t grow it you haven’t explained it” became the credo of agent-based simulation most prominently introduced by Epstein and Axtell in their seminal book “Growing Artificial Societies”. The book at hand agrees with this credo in principle, and its subtitle (A Toolkit) promises to fix some of the potholes standard economic theory has created over the years.Chapter 4 thus establishes the background against which the necessity for agent-based models in economics becomes apparent: the struggle of standard macroeconomic theory to accommodate boundedly rational agents that are heterogeneous in their preferences and expectations. However, the chapter offers little guidance on how to use agent-based approaches in macroeconomics. It merely suggests resorting to decision heuristics and learning models that allow agents to adopt behavioural rules that maximise their utility in a given context and refers the reader to the next chapter for details. Chapter 5 however, after a reiteration of the “shortcomings” of homo oeconomicus and a brief review of the literature on bounded rationality, prospect theory and heuristics, jumps straight into models of learning without connecting these strands of literature or providing clear guidelines on how to use these concepts in agent-based modelling. Chapter 6 then deals with agent interactions and gives a proper review of the topics that enthuse many agent-based modellers in the social sciences: micro-macro models and social networks.
The third part of the book, with chapters 7 to 9, deals with agent-based methodology. Although the three chapters address different topics (experimentation, validation and estimation, respectively), all three chapters reveal interesting facts about the conditions under which agent-based models can tell us something about the real world. Maybe most important is the treatment of ergodicity of stochastic processes. Simply put, if a stochastic process is ergodic, initial conditions have no bearing on the outcome of the process in the long run. In contrast, the outcome of a non-ergodic stochastic process depends crucially on initial conditions. Hence, if initial conditions are unknown and the process takes place only once (e.g., the events that led to Google being the dominant search engine on the internet), then trying to explain the outcome using agent-based simulation experiments is less meaningful. This does not mean however that non-ergodic processes cannot be modelled and studied; if the occurrence of a particular process is frequent (e.g., the emergence of behavioural regularities in lab experiments), agent-based models can be used to derive hypotheses about initial conditions. These hypotheses, in turn, can be tested and the models thus empirically validated.
Although the book is not more than the sum of its parts, there is something important to be learned, in particular from the last three chapters: The book sketches the boundary conditions of agent-based methodology and puts a different complexion on the credo “If you can’t grow it you haven’t explained it”.
Return to Contents
of this issue