Split-Gate Memtransistors for Energy-Efficient Adaptive Reinforcement Learning

Real-time, on-chip learning has become increasingly important as artificial intelligence edge-computing systems are deployed in dynamic environments ranging from autonomous vehicles to robotics, where static pretrained models are insufficient. In this context, analog in-memory computing hardware has been explored to alleviate energy and latency bottlenecks, but its limited reconfigurability has hindered the implementation of reinforcement learning in deployed artificial intelligence agents. Here, we demonstrate split-gate MoS 2 memtransistors where local field-effect gating of the Schottky contacts and semiconducting channel enables improved control of memristive switching ratios and conductance states, respectively. These characteristics enable efficient implementation of reinforcement learning due to the combination of nonvolatile synaptic weight updates and rapid parameter adjustments. Adaptive reinforcement learning performance is benchmarked with a cartpole balancing task that serves as an elementary robotic example for embodied decisions and actions, ultimately showing a 6-fold improvement in total reward and a 5-fold reduction in the number of programming steps.