Abstract:
Adaptive neuro-fuzzy inference system (ANFIS) is useful in predicting outcomes of event,
process or system with its effectiveness relying on some factors most especially its set of rules
that models data behaviour, and parameters (antecedent, consequent and rule base). Several
promising meta-heuristic optimization techniques like Ant Colony, Genetic Algorithm, Particle
Swarm Optimization (PSO), and Invasive Weed Optimization have been proposed to improve
on the performance of this model. However, majority of these techniques improved ANFIS
prediction accuracy through the optimization of its antecedent and consequent parameters
while little or no attention was drawn to its rules. Also, absolutely no attention has been drawn
to the quality of rules extracted by an approach (grid partitioning, subtractive clustering or
fuzzy c-means) employed by ANFIS as often implemented in popular statistical packages like
MATLAB and R. In this study, an improved ANFIS-based Predictive Framework that uses a
probability trajectory and k-nearest neighbour-based clustering ensemble of fuzzy c-means and
PSO to optimize the quality of its rules and parameters (antecedent and consequent)
respectively was developed. MATLAB 2018b was used as the development environment for
the predictive framework. From the experimental results on lithology prediction from two oil
wells characterized by Dataset1 and Dataset2 (petrophysical datasets or raw oil well logs
collected from Niger Delta region of Nigeria) as presented in the thesis where mean squared
error (MSE) was used as objective function, shows that ANFIS (at least with its MSE less than 0.5 in all cases) could provide a good lithology prediction from both wells. However, when
ANFIS was optimized by probability trajectory-based clustering ensemble, it could provide a
better prediction even when the k-nearest neighbourhood of it was set to a value from 5, 10, 15
and 20. Also, the stability (in terms of variation in MSE) of ANFIS and the improved ANFIS
for different executions (in each of the executions, the systems considered comparatively
twenty rounds of holdout cross-validation where, each dataset was partitioned into 70% training dataset and 30% test datasets), better than that of ANFIS and any of its previous
improvements.
