We propose a novel tag selection algorithm to improve the efficiency of tag arbitration in RFID systems. Based on the probability that a given tag is located in the sensing range of the reader, our algorithm involves two steps. First, tags with higher probabilities are selected and directly arbitrated by their IDs, in order to eliminate the collisions caused by multiple tag replies and reduce the number of empty tag replies. And then, the remaining unidentified tags, which will be in a small quantity if any, are arbitrated with the Query-Tree or the frame-slotted Aloha protocols specified in the RFID standards. Our proposed Probability-based Tag Selection (PTS) algorithm is compatible with the RFID standards with very minor modification on the readers and no changes on the tags. An analytical model is constructed to find the optimal threshold of tag selection in the first step. The simulation shows the result from our model nicely represents the optimal threshold value. It further demonstrates that PTS significantly reduces the arbitration delay and improves the arbitration rate with a given budget, comparing with the Query-Tree and frame-slotted Aloha protocols.