Low-power Autonomous Adaptation System with Deep Reinforcement Learning

Abstract

Recently, a lot of autonomous systems such as self-driving have been as close to human-level due to the rapid improvement of deep neural networks (DNN). However, they only performed well in the pre-trained environment and cannot achieve the same performance for the sudden environmental changes. To resolve this issue, self-adaptation through deep reinforcement learning (DRL) has been highlighted. However, it is hard to utilize DRL in the autonomous system due to the massive memory bandwidth requirements. In this paper, we propose a low-power and high-performance DRL system with an energy-efficient DRL chip. The proposed DRL chip can seamlessly compress both weight and feature map to reduce the number of memory access. The proposed system with DRL chip demonstrates adaptation of humanoid to sudden environmental changes at Mujoco Humanoid-v2. The proposed system shows 10.3 iteration/J of training energy efficiency which is 3.9× higher than NVIDIA TX2.