This paper investigates the performance of particle swarm optimization (PSO) and unified particle swarm optimization (UPSO) in magnetoencephalography (MEG) problems. For this purpose, two interesting tasks are considered. The first is the source localisation problem, also called the 'inverse MEG problem', where an unknown excitation source has to be identified, based on a set of sensor measurements that can be contaminated by noise. We refer to the second task as 'forward task for inverse use'. It consists of the detection of the proper coefficients for approximating the magnetic potential through a spherical expansion, as accurately as possible. Also, the study of their behaviour under variations of the number of available measurements is considered. The obtained results are statistically analysed, providing useful insight regarding the applicability of the employed algorithms on such problems. Also, significant indications regarding the behaviour of several intrinsic dependencies of the problem are derived.