[20a-F211-1~10] 31.1 フォーカストセッション「AIエレクトロニクス」

▼ [20a-F211-8] In-memory Reinforcement Learning Hardware with Stochastic Conductance Change of Ferroelectric Tunnel Junctions

〇(P)Radu M Berdan¹、Takao Marukame¹、Kensuke Ota²、Masumi Saitoh²、Shosuke Fujii²、Jun Deguchi²、Yoshifumi Nishi¹ (1.Toshiba RDC、2.Toshiba Memory)

キーワード：memristor, neural networks, FTJ

Building compact and efficient reinforcement learning (RL) systems for mobile deployment requires departure from the von-Neumann computing architecture and embracing novel in-memory computing, and local learning paradigms. We exploit nano-scale ferroelectric tunnel junction (FTJ) memristors with inherent analogue stochastic switching arranged in selector-less crossbars to demonstrate an analogue in-memory RL system. That is, via a hardware-friendly algorithm, capable of learning behavior policies. We show that commonly undesirable stochastic conductance switching is actually, in moderation, a beneficial property which promotes policy finding via a process akin to random search. We experimentally demonstrate path-finding based on reinforcement (Fig. 2), and solve a standard control problem of balancing a pole on a cart via simulation, outperforming similar deterministic RL systems.