Modifying the velocity in adaptive PSO to improve optimisation performance

This article investigates the evolution of the velocity vector as the AdPSO (Adaptive PSO) algorithm optimizes a set of parameters through a number of epochs. Experimental results have shown that when using a swarm to find the optimal solution to a specific natural language processing (NLP) application gradually the velocity vector is decreased towards a very small value that causes the particles to switch from exploration (i.e., the attempt to determine radically new solutions) towards exploitation (search of solutions that are close to those already identified). Based on this observation, a study is carried out to determine whether the velocity vector may be handled in a more efficient manner. An algorithm for reinitializing the velocity of swarm particles is proposed, which improves the exploration of the swarm, by reenergizing particles that have very low velocities. Also, the effect of bounding the initial velocity of particles is studied, to determine whether improved optimization performance can be achieved. The effectiveness of the velocity reinitialisation mechanism is further examined by application to a selection of benchmark test functions. These experimental results are supplemented by relevant statistical tests that indicate a significant improvement in many cases.