Fully Tunable In-Memory Eligibility Traces Based on Ferroelectric Semiconductor Field-Effect Transistors

Traditional implementations of eligibility traces have struggled with the tunability due to fixed relaxation characteristics, usually necessitating energy-intensive auxiliary circuitry. In this work, we introduce a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{Bi}_{2} \mathrm{O}_{2} \text{Se}$</tex> ferroelectric-semiconductor field-effect transistor-based reinforcement learning system that offers a fully tunable eligibility trace ranging from 0.001 to 0.999, spanning three orders of magnitude. Benefiting from the ferroelectric properties of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{Bi}_{2} \mathrm{O}_{2} \text{Se}$</tex>, our device functions as a non-volatile 7-bit memory, featuring an impressive memory window of 10.5 V. In a demonstration of the Taxi Problem, our system achieves an average reward of 8.39 over 100 epochs while realizing a remarkable 99.68 % reduction in energy consumption compared to conventional GPU implementations.