DynTLBO - A Teaching Learning-based Dynamic Optimization Algorithm

Abstract

The Teaching Learning Based Optimization (TLBO) algorithm simulates the knowledge-transfer process between teacher and learners as well as between peer learners. Although TLBO has been already successfully applied to both constrained and unconstrained engineering optimization problems, it sometimes prematurely converges toward local optima, especially in high dimensional, multimodal, or deceptive fitness landscapes. We therefore propose to further characterize the limitations of TLBO by investigating its performance on different benchmarks featuring both stationary, to establish a baseline, but especially non-stationary fitness landscapes. The results are then compared with a state of the art population-based optimization algorithm (Differential Evolution - DE) and its variants Self Adaptive Differential Evolution (jDE) in order to establish the suitability of TLBO on such landscapes. We found that TLBO exhibits a pronounced imbalance in its exploration vs. exploitation tradeoff that prevents it from maintaining a diversified population. It is well known that maintaining diversity as the population converges, and more generally balancing the exploration versus exploitation tradeoff, are both essential considerations in any population-based optimization technique. This is especially true for non-trivial problems where premature convergence to local optima is more likely. We therefore also proposed a novel TLBO variant that better manages population diversity and is therefore more suitable for dynamic optimization applications. We found that the resulting DynTLBO algorithm showed significant performance improvements on commonly used benchmarks.