The paper considers the problem of how a distributed system
of agents (who communicate only via a localised network) might achieve consensus
by copying beliefs (copy) from each other and doing some belief pruning
themselves (drop). This is explored using a social simulation model, where
beliefs interact with each other via a compatibility function, which assigns a
level of compatibility (which is a sort of weak consistency) to a set of
beliefs. The probability of copy and drop processes occurring is based on the
increase in compatibility this process might result in. This allows for a
process of collective consensus building whilst allowing for temporarily
incompatible beliefs to be held by an agent.
This is an example of socially-inspired computing (by analogy with biologically-inspired). The space of behaviours in a MAS where agents interact with each other at the same time as reasoning/learning themselves is so vast that a "structuring idea" is needed. Here we apply an analogy with human opinion-dynamics as an analogy with which to design, manage and understand a subset of this huge space. To make this more precise this idea is instantiated in a social simulation which can then be used as a proto-theory about the MAS themselves. Such simulations can be used as part of a chain of increasingly realistic simulations, each simulation acting as the "theory" for the more detailed one, thus allowing a cautious approach to "designing" MAS with emergent properties. The model here can be seen as part of a series of "opinion dynamics" models that have been published, but departs radically in having beliefs represented and interacting, rather than artificial "susceptibilities" etc. It is thus a much more appropriate model for studying opinion dynamics in MAS.
Results suggest that the rate and effectiveness of consensus building in such systems is complex, the belief structure (i.e. the compatibility of beliefs); the network topology; the copy and drop rates; and whether agents only drop beliefs if they are inconsistent (compatibility < 0); all affect the rate of consensus building. On the whole a balance between the things that cause variation and the speed of belief "convergence" need to be balanced. The results are compared to those from the simple opinion dynamics models, and some tentative hypotheses about agent consensus in this model posited.