Meta-Learning Architectures for Adaptive Real-Time Decision-Making in Non-Stationary Environments

Abstract

Decision making in dynamic and non-stationary environments is a challenging problem for today's autonomous systems, which is largely because of the high computation burden and catastrophic forgetting of the conventional machine learning approaches. The current study explores the physical implementation of an adaptive meta-learning architecture specifically designed for edge computing platforms with ultra-low latency and limited resources. The proposed architecture combines two key ideas: first-order Model-Agnostic Meta-Learning (MAML) with Meta-Policy Gradient Reinforcement Learning (MPG-RL) to achieve rapid adaptation to few-shot tasks and actively combat catastrophic forgetting across sequential data streams. Empirical testing was performed on real physical testbeds with ARM Cortex-M microcontrollers and Jetson Orin edge devices, in order to strictly quantify performance in real-world non-stationary drift conditions, without resorting to the use of a hypothetical testbed. The performance of the proposed architecture was exhaustively evaluated under three key parameters, namely task adaptation latency, energy consumption, and accuracy of inferences retained after adaptation. The empirical results show that the adaptive meta-learning framework can reduce task adaptation latency and can achieve an energy efficiency improvement of up to 50% compared to the standard deep reinforcement learning baselines. The research offers a scalable, fully human-independent solution to next-generation edge intelligence with transformative benefits for autonomous robotics, industrial internet of things and intelligent transportation systems.