The operation of high frequency devices can be severely deteriorated due to the tolerances introduced by manufacturing, especially when more than one manufacturing process is required in fabrication of the device. The effects of parameter deviations can be mitigated at the design phase through optimization of the yield, but this process is very expensive, as each iteration calls for a full yield analysis. This paper will discuss swarm intelligence parametric optimization for improved yield of a Ka-band integrated filter-antenna, using NLPLS-PCE. This technique has shown to provide an excellent method for calculating yield for microwave devices and antennas [1, 2]. The technique creates a reduced set of latent variables, which is then approximated through the PCE, by a set of polynomials which are orthogonal under the weighting of a distribution function. In [2], it was shown that the yield of a 93-variable Square Kilometre Array (SKA) Mid-Frequency Aperture Array (MFAA) Log Periodic Dipole Antenna (LPDA), where all dimensional variables are assumed to have a Gaussian distribution function, could be accurately estimated using only 80 full 3D electromagnetic analysis frequency sweeps, instead of hundreds required by a Monte-Carlo analysis. The integrated filter-antenna consists of a ridge-guide filter terminated in a Vivaldi antenna. The yield specification for this problem includes that the reflection coefficient must be below -15 dB across the band. The number of statistically independent dimensional variables is 19. Preliminary yield analysis shows that only 30 frequency sweeps are required to accurately estimate yield of the Ka band filter.