|
Abstract:
|
"In this work we investigate learning as performed by agents interacting as they infer models of a complex system representable by a complex network, under the presence of observation errors. The models correspond to estimations of the adjacency matrix of the complex system under investigation. We focus the specific case in which, at each time step, each agent takes into account its just performed observation as well as the average of the models of its neighbors. A series of interesting results are identified with respect for Barab\'asi-Albert interaction networks. First, it is shown that the interaction among agents allows an overall improvement in the quality of the estimated models, a consequence of the averaging among neighbors. We then investigate situations in which one of the agents has different probability of observation error (twice as much higher or lower than the other agents). It is shown that the influence of this special agent over the quality of the models throughout the rest of the network is substantial and varies linearly with the respective degree of the agent with different estimation errors. In case the degree of each agent is taken as a fitness parameter, in the sense that the influence of the node over the other agents is proportional to its degree, the effect of the different estimation error is even more pronounced, becoming superlinear."
|