Reinforcement Learning System Comprising Resistive Analog Neuromorphic Devices

Reinforcement learning algorithms are widely used in many practical artificial-intelligence (AI) applications. Herein, we propose a compact hardware-based learning system comprising analogue resistive memory devices called “resistive analogue neuromorphic devices (RAND).” We begin by showing that the resistance of a RAND linearly varies with the voltage pulses, meaning that a single RAND is sufficient to solve a reinforcement learning problem (i.e. the “2-armed bandit problem”) using “tug-of-war (TOW) dynamics.” Next, we show that 2
^k
-armed bandit problems are also solved by hierarchically combining 2
^k-1
RANDs. Finally, we numerically demonstrate that the proposed methods are promising as compared with conventional methods.