In general, conventional Arbiter-based Physically Unclonable Functions (PUFs) generate responses with low unpredictability. The N-XOR Arbiter PUF, proposed in 2007, is a well-known technique for improving this unpredictability. In this paper, we propose a novel design for Arbiter PUF, called Double Arbiter PUF, to enhance the unpredictability on field programmable gate arrays (FPGAs), and we compare our design to conventional N-XOR Arbiter PUFs. One metric for judging the unpredictability of responses is to measure their tolerance to machine-learning attacks. Although our previous work showed the superiority of Double Arbiter PUFs regarding unpredictability, its details were not clarified. We evaluate the dependency on the number of training samples for machine learning, and we discuss the reason why Double Arbiter PUFs are more tolerant than the N-XOR Arbiter PUFs by evaluating intrachip variation. Further, the conventional Arbiter PUFs and proposed Double Arbiter PUFs are evaluated according to other metrics, namely, their uniqueness, randomness, and steadiness. We demonstrate that 3-1 Double Arbiter PUF archives the best performance overall.