Simulation Of Coalition Formation With Heterogeneous Agents By Swarm

This paper analyzes the process of social aggregation in an environment in which there are agents with different endowments that can form coalition to produce and divide an output; there is an aggregation force determined by an increasing returns to scale technology, but imperfect information on other agents' action does not allow for the formation of an unique coalition in which all agents participate. In this setting we examine how the outcomes of social interaction depend on the basic characteristics of the economy, i.e. the returns to scale of the available technology that converts individual actions in coalitional results and the distribution of resources. The dynamics of model allows for a first stage where every agent sends signals (messages) to the other agents, consisting in non binding proposals to coalesce, to verify which agents is going to coalesce with him and therefore to coordinate their actions; we can see this as a pre-game communication phase, where to send a signal does not imply any cost (also in terms of future agreements). The outcome of this process of learning and coordination is the formation of a coalitional structure and the resulting output is divided among the members of coalition. We assume that agents elaborate their signals by genetic algorithms, that seems particularly suited in this contest where agents have bounded rationality. The use of genetic algorithms in elaborating agents' signals is motivated by important contributions from the theory of cognitive processes; in this approach, agents, when called upon to make a choice in a complex environment, do not make explicit optimization, but rather operate on a limited set of rules (or mental schemes) that they continuously modify reacting to the effects of their own behaviour. The crucial point of this adaptive process is how new rules are formed and, from this point of view, genetic algorithms offer a very intuitive method inspired to the process of natural selection: the good rules not only persist in the set of the individual rules, but they provide the basis for the creation of new ones, through their recombination; this is the central feature of the learning process. The analysis of the model's properties is mainly performed by simulation, focusing on the effects of variations of technology and of initial distribution of resources. The simulation is performed in the Swarm environment, which is very efficient in the simulations of artificial life structure and suits well the present case of an artificial economy in which there are many heterogeneous interacting agents. In our particular case, the most useful features of the Swarm environment is the possibility to model a prototype agent (essentially consisting in specifying its decision process) and to generate the population of heterogeneous agents by replicating this agent, changing its characteristics (see its initial endowment), while the interaction among agents is managed by Swarm. A robust result seems to be that an increase in inequality of initial endowments, given a certain technology, tends to reduce average payoffs, that is the economic efficiency; there is a suggestive similarity between this finding and those of the literature on growth and distribution. Moreover the process of coalition formation tends to increase the initial inequality, so that in a dynamic perspective this aspect appears particularly relevant. Other interesting point is the effects of change in technology. Even if it is not possible to reach an unambiguous conclusions, we find that for the most cases an increase in returns to scale causes a higher inequality; in particular, for some distributions it seems to be a common aspect that the coalitional structure is polarized, that is the richest agents coalesce only among themselves, so as the poorest ones.