TR-14-07: Evolving Neurocontrollers for the Control of Information Diffusion in Social Networks

Abstract

Automated control of information diffusion in social networks is a difficult problem with potential application to marketing, security, and social organization. The θ-Consensus Avoidance Problem (θ-CAP) is a social network control problem modelling the specific case of maintaining a network within a homoeostatic range of states. It is an important problem because it represents social networks avoiding extreme views or situations that might lead to extreme behaviour. This paper presents a comparison of two Evolutionary Artificial Neural Network (EANN) variants acting in the behavioural component of an autonomous control system for instances of the θ-CAP. A novel variant of EANN is proposed by adopting characteristics of a well-performing heuristic into the structural bias of the neurocontroller. Information theoretic landscape measures are used to analyse the problem space as well as variants of the EANN. The results obtained indicate that the two neurocontroller variants learn very distinct strategies for balancing the states of the social network, however, the newly proposed variant demonstrates improvements in both solution quality and execu- tion time. A ramped-difficulty evolution scheme using constraint parameters is demonstrated to be effective at creating higher quality results as compared to the standard scheme for EANNs. A correlation between the proposed instance difficulty and identifiable landscape characteristics is discovered as well.

TR-14-07.pdf