Reviewed by
Ermanno Catullo
Department of Economics and Social Sciences, Politechnic University of Marche, Ancona
Artificial economics offers a plurality of methodological instruments for dealing with economic system complexity. As pointed out in the introduction, the papers presented in Managing Market Complexity: The approach of Artificial Economics interpret economic phenomena as emerging patterns of complex adaptive systems. Agent based modelling is the main approach followed by these papers to deal with complex economic systems. Indeed, agent based simulations allow to model agents heterogeneity and their decentralized interaction, which may lead to the emergence of macro and micro patterns (Gilbert and Terna 2000). Artificial economics underlines the flexibility of agent based modelling methodology which may be applied in several economic fields at different levels of abstraction: networks, macroeconomics, finance, industrial organization and management. Although these papers share the same ontological and methodological assumptions, they present considerable differences in terms of agents behaviour and interaction. Thus, in my review, I will try to briefly report some of their common characteristics.Let’s start by considering that agents can be different according to their learning capabilities. Firstly, they can make different choices using algorithms with different learning capabilities. Secondly, each decision heavily depends on the information set on which is based. If the information set on which choice is made is restrict and incomplete, even a maximizing choice may be relatively effective. Third, complex environments have a certain degree of uncertainty that reduces the forecasting capacities of agents. Adding some stochastic component in decision making could help to understand individual choices that is hard to model exhaustively.
For instance, in the Diedrich and Beltran paper on Internet traffic discrimination policies, maximization procedures are used by consumers to value provider supply. However, due to bounded rationality, consumers that want to change provider explore only a limited set of providers until they reach a satisfying solution. In the paper of Cruciani and colleagues on sense making and cooperation, agents are maximizers but their choice capabilities are limited by a restricted information access.
Agents follow heuristics in many works. For instance, in the work of Provenzano on contagion and bank run, the probability that a depositor withdraws, which potentially may trigger a bank run, is given by a simple rule. Even in the Erlingsson and colleagues’ paper, which integrates housing market in a macro model, consumer can sell or buy an house randomly but with notable exceptions in which they can decide deterministically.
Models of learning helps to consider more elaborated cognitive capabilities. For example, in the Guerci and Rastegar’s model, which compares two specifications of a realistic electricity market, agents choose their bids following a reinforcement learning algorithm inspired by Roth and Erev (1995). The importance of learning capabilities is also underlined in Kopanyi’s work, where the authors replicates a Bertrand competition market with differentiated goods, where agents may switch among competing learning rules according to their perceived effectiveness. The switching between learning rules may assume cyclical patterns, as the adoption of a particular learning rules dynamically affects the aggregate results and, consequently, has an impact on its effectiveness and on the effectiveness of the other rules.
Furthermore, agents cognitive capacities may be strongly influenced by the agent interaction structure. For instance, the model of Osinga and colleagues starts from the idea that information about market prices about product is not only conveyed by price but is related to the network of interactions as this can influence the diffusion of information on demand and supply. Mostly, the interaction structure is explicitly defined as networks between linked agents. For instance, Provenzano’s work shows bank network configurations that are crucial to determine the resilience of the credit system, so that different specifications of the credit network are examined. Similarly, different network configurations are analyzed in a paper by LiCalzi and Milone, where simple organizational architectures are compared that are characterized by the relation between hierarchical order and information flows.
Interestingly, there are examples where the interaction structure empirically calibrated. For instance, Guerci and Rastegar examine how electricity flows run on a grid that reproduce exactly the zonal market structure and the relative transmission capacity between neighbouring zones of the Italian grid model. A more elaborate structure of the interaction network is presented by Schouten and colleagues, who investigated rural landscapes. Their model defines specific spatial landscapes to reproduce agents heterogeneity in economic and environmental terms so that it can be initialized using empirical data of farms and agricultural spatial structure and describe the co-evolution of economic agents and landscape in which they operate. Similarly, Osinga and colleagues examine market supply in agriculture by reproducing crucial aspect of real environment and of network relationships among farmers. Here, information circulates in network of family members, neighbourhood and friends. While the first two are static networks, the third one is dynamic, i.e., friendship relations change during the simulation affecting information diffusion. This means that network configuration dynamics can shape the relationship between macro and micro dimensions of the economic system. The importance of network evolution is crucial in Lopolito and colleagues’ model, where new technologies spread from niches. In this case, the strength of a new technology depends on the emergence of a stable network with a sufficient critical mass of adopter to sustain the diffusion of the innovation. Furthermore, Cruciani and colleagues examine cooperation by discussing the influence of agents’ heterogeneity through homophily on the dynamic network configurations.
Although the papers included in this book touch upon different cognitive capabilities and look at varying interaction mechanisms, they try to determine the robustness and validity of simulation results. They test the robustness of the model via sensibility analysis, varying parameters values or particular specifications. Moreover, they establish the robustness and the external validity of their result through parameter calibration and reproducing empirical evidences. Indeed, as underlined by Diks and Makarwicz’s paper, parameter configuration and especially initialization procedures are crucial in determining the dynamics of the simulations. They propose a genetic algorithm procedure to parameter estimation. Similarly Grazzini and colleagues look at effective estimation procedures for agent based models, testing them on an agent based implementation of the well-known Bass innovation diffusion model.
To sum up, the book papers included in this collection provide a brilliant overview of the advances that artificial economics has made in analyzing complex economic systems. Although the models presented are diverse both in terms of substantive focus and research aims, these papers share a common ontological and methodological background, which can help us to understand prospects and limitations of the agent-based approach to economics. The possibility of exploration opened by the peculiar flexibility of agent based modelling requires intelligent adaptation to different empirical scenarios. Moreover, agent-based modelling helps us to exploit counterfactual experiments, which can shed new light on important economic phenomena and prearrange theoretical models to empirical validation. Once having said this, we must also say that, in order to pave the way for embedding agent-based models in economics, improving the statistical robustness of simulation results, validating models with empirical data and favouring findings’ replicability are challenges that must be addressed more seriously in the future.
ReferencesROTH, A. E. and Erev, I. (1995). Learning in extensive form games: Experimental data and simple dynamic models in the intermediate term. Games and Economic Behavior, 8, pp. 164-212.
Return to Contents
of this issue