Microwave Kinetic Inductance Detectors (MKIDs) are extremely sensitive radiation detectors based on superconducting resonators that can be combined in large arrays on a single readout line within a limited frequency bandwidth. This makes MKIDs ideal detectors for the ultimate far-infrared observatory: a future space-based actively cooled telescope with its performance solely limited by the low universe background radiation. However, to reach these detector requirements, state-of-the-art MKIDs still need a order of magnitude improvement in device sensitivity. In this work, the MKID sensitivity is improved by reducing the aluminium volume that absorbs pair-breaking radiation into quasiparticle excitations, while making sure all radiation is still absorbed. Furthermore, a key requirement is sufficient reduction of excess noise as to keep the device intrinsically limited by thermally driven random fluctuations in the number of quasiparticles in absence of radiation, or Generation-Recombination (G-R) noise. To this end, a model is developed that describes the noise contributions as function of device geometry, readout power, material properties and radiation power. Subsequently, a realistic MKID design is presented and tested that reduces excess noise and maximises the sensitivity, expressed as Noise Equivalent Power (NEP). At high temperatures, good overall agreement is found between the measured noise spectra and the model. At low temperature T = 120 mK, the measurement results give an optical NEP similar to current state-of-the-art MKIDs. The NEP is not as low as expected due to short quasiparticle lifetimes, an unexpected decrease in the G-R noise level and a very high excess noise attributed to Two-Level Systems (TLS) noise that starts to dominate the already low G-R noise spectrum at low temperatures. Possibly, the quick quasiparticle lifetime saturation and noise level drop are caused by a strong readout power effect, as the readout power is known to create excess quasiparticles and to cause a strongly non-thermal electron energy distribution in the aluminium strip of the MKID. However, the exact microscopic details of these effects are unknown and not studied in this project. Based on the current chip design, a straightforward way to improve device performance and study the readout power effect in more detail is a reduction of the high TLS noise levels, which is possibly fabrication related. This would allow an unobstructed view of the G-R noise spectrum at low temperatures, thereby allowing both a study of the readout power effect on the quasiparticle system, and ultimately achieving the improvement in NEP needed reach the detector requirements for the ultimate space-based far-infrared observatory.