Loading...

An Agent-Based Modeling of Consensus in Complex Networks with Machine Learning Methods

Momeni Taramsari, Naghmeh | 2012

564 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 43764 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Golestani, Jamaloddin; Ghodsi, Mohammad; Nobakhti, Amin
  7. Abstract:
  8. In this thesis, the individuals’ decision process to achieve consensus in complex networks is modeled. A set of experiments with volunteered participants were run, in which the person is a node in a network with 35 other nodes who behave based on a prescribed model. Each participant is tested twice, once with and once without reward. Performance of individuals in the two experiment types and the effect of structural properties of the underlying network on the results are studied. It is observed that increase in the characteristic path length, diameter, clustering coefficient and local efficiency of the network adversely affects consensus.On the other hand, global efficiency and closeness centrality of the participant facilitates consensus. Also, a number of individual behavioral factors are defined and calculated for participants. To model the empirical data, machine learning methods are employed. Data are modeled using logistic regression, naïve Bayes, K nearest neighbors and decision trees.The latter three produce an average error of 0.14, outperforming logistic regression with an average error of 0.21. At the end, the dynamics is implemented on a large Barabasi-Albert and a large Watts-Strogatz network, as well as on the initial graph used for the experiments.We run these simulations using three decision schemes for the nodes: the behavioral model acquired by the decision tree method, the initial model used in the experiment, and a simple follow-he-majority model. In the 36-node network and the Watts-Strogatz graphs, the decision tree model has the best performance with 88% and 74% consensus, respectively.On the Barabasi-Albert graph, the initial model designed for the experiment has the best performance with 97% consensus.
  9. Keywords:
  10. Complex Network ; Machine Learning ; Agent Based Modeling ; Consensus

 Digital Object List

 Bookmark

No TOC