A computational intelligence approach to modelling interstate conflict : Forecasting and causal interpretations

Abstract

The quantitative study of conflict management is concerned with finding models which are accurate and also capable of providing a causal interpretation of results. This dissertation applies computational intelligence methods to study interstate disputes. Both multilayer perceptron neural networks and Takagi-Sugeno neuro-fuzzy models are used to model interstate interactions. The multilayer perceptron neural network is trained in the Bayesian framework, using the Hybrid Monte Carlo method to sample from the posterior probabilities. It is found that the network is able to forecast conflict with an accuracy of 77.3%. A hybrid machine learning method using the neural network and the genetic algorithm is then presented as a method of suggesting how conflict can be brought under control. The automatic relevance determination approach and the sensitivity analysis are used as methods of extracting causal information from the neural network. The Takagi-Sugeno neuro-fuzzy model is optimised, using the Gustafson-Kessel clustering algorithm to partion the input space. It is found that the neuro-fuzzy model predicts conflict with an accuracy of 80.1%. The neuro-fuzzy model is also incorporated into the hybrid machine learning method to suggest how the identified conflict cases can be avoided. The casual interpretation is then formulated by a linguistic approximation of the fuzzy rules extracted from the neuro-fuzzy model. The major finding in this work is that the interpretations drawn from both the neural network and the neuro-fuzzy model are consistent.