Both pulse shape discrimination methods based on temporal difference characteristic and phoswich detectors are well-known and effective for particles discrimination, so the combination and further improvement are significant for better performance. In this paper, a phoswich detector combining YAG(Ce) and CsI(Tl) was constructed for α/β/γ discrimination using various techniques including two machine learning algorithms and an improved charge comparison method that further exploited the difference in the decay rate of pulse tail. The detector was designed with Geant4, simulating the suitable sizes of scintillators, and a high-speed digitizer was used for the collection of pulses that were from three radioactive sources. The results among the α (β) and γ rays had the figure of merit values greater than 2, simultaneously with an optimal 1.68 of the improved method that had more than twice the performance comparing the original 0.69 for α/β identification, which showed the detector is able to discriminate α, β and γ rays in mixed radiation field effectively.