Neuromorphic computing offers memory collocated computing architectures that are brain-inspired and inherently supports on-device learning capabilities. Ferroelectric field-effect-transistor (FeFET) has been demonstrated as a multi-state neuromorphic synaptic device. The analog FeFET synapse exhibits varying synaptic weights in response to different write programming schemes. In this research, we explore the impact of different write programming schemes on the accuracy of a neuromorphic system. A simple write programming scheme with saw-tooth pulse is proposed. The benchmark algorithm is a rate based multilayer neural network that employs feedback alignment backpropagation for training. The performance of the proposed schemes is demonstrated on three different datasets.