Reinforcement Learning in Memristive Spiking Neural Networks through Modulation of ReSuMe

摘要

In this paper, a novel hardware-friendly reinforcement learning algorithm based on memristive spiking neural networks (MSNN-RL) is proposed. Neurons for spike coding are designed specifically to complete transformation between analog data and discrete spikes. Then, remote supervised method (ReSuMe) is used to combine SNN with basic reforcement learing (Sarsa). Besides, bionic memristive snynapses are designed to speed up ReSuMe. Furthermore, the circuit scheme of MSNN-RL is designed with modulation of memristor synapses. Finally, the application of MSNN-RL in acrobot system is discussed. Simulation results and analysis verify the effectiveness of the proposed algorithm (MSNN-RL) and show it is superior to traditional apporach.