A research team led by Professor Yoo Ho-cheon from the Department of Electronic Engineering at Hanyang University, in collaboration with Professor Shin Won-jun of Sungkyunkwan University and Professor Amit R. Trivedi of the University of Illinois Chicago, announced on the 17th that they have developed a Gaussian-Sigmoid Reinforcement Transistor (GS-RT) that can implement both Gaussian and Sigmoid activation functions within a single device—simply by adjusting the gate voltage.

This transistor features a structure that allows free switching between activation functions within a single device, providing a hardware-level solution to the critical exploration–exploitation trade-off in reinforcement learning (RL) algorithms. The team expects the GS-RT to significantly simplify RL-based AI circuits while enhancing both learning efficiency and overall performance.

Traditionally, RL systems rely on both Gaussian and Sigmoid 토토사이트 은행 조회서. However, hardware implementation of these 토토사이트 은행 조회서 typically requires complex calculations and dozens of transistors, resulting in excessive power consumption and design complexity. Circuits demanding both 토토사이트 은행 조회서 also increase chip area and resource usage, posing major obstacles to efficient AI semiconductor design.

To overcome these challenges, the team constructed a heterojunction structure combining an oxide semiconductor (IGZO) and an organic semiconductor (DNTT) in a vertically stacked n–p–i–p formation, integrating an insulating layer and asymmetric electrodes. This configuration enables the device to exhibit either Sigmoid or Gaussian current characteristics through gate voltage control alone, creating a structure optimized for neuromorphic computing that can replace existing complex circuits with a single device.

The GS-RT device demonstrated practical effectiveness in reinforcement learning experiments. In the multi-armed bandit problem, it achieved approximately 20% faster learning and 30% higher rewards compared to conventional methods. In the cart-pole experiment, the GS-RT outperformed the standard ReLU function in terms of convergence speed and showed high stability even under noisy conditions. These improvements are attributed to the GS-RT’s nonlinear characteristics, which provide stable gradient flow during training.

“This research presents a new possibility in AI semiconductor technology by enabling dual-mode activation functions within a single device using only gate voltage control,” said Professor Yoo Ho-chun, the project’s principal investigator. “It is expected to serve as an efficient hardware solution for next-generation AI semiconductors, particularly in reinforcement learning, neuromorphic systems, and ultra-low-power edge computing.”

The 토토사이트 은행 조회서 was supported by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation (IITP) through its AI Semiconductor Talent Cultivation Program, as well as by the National 토토사이트 은행 조회서 Foundation of Korea (NRF). The study was published in the July 2025 issue of Advanced Functional Materials (IF: 19.0, top 5% in JCR).

The paper, titled “Gaussian-Sigmoid Reinforcement Transistors: Resolving Exploration-Exploitation Trade-Off Through Gate Voltage-Controlled Activation Functions”, lists master’s student Park Ji-su as the first author and doctoral student Seo Ju-hyeong as co–first author, with Professor Yoo Yoo Ho-cheon as corresponding author.