Adaptive Random Number Generator Based on RRAM Intrinsic Fluctuation for Reinforcement Learning

In this work, a novel random number generator with adjustable adaptability is demonstrated based on the intrinsic fluctuation of resistive random access memory (RRAM). Experimental data shows the standard deviation of resistance state has a negative dependence on compliance current (CC), thus an adaptive random number generator (ARNG) can be realized by controlling the CC in RRAM set process. Benchmarking results show that the RRAM based ARNG can achieve same performance compared with software ARNG in reinforcement learning process of zebra finch's singing pattern.