A GA-RL Hybrid Framework for E-Paper Waveform Optimization

Ziwei Mo, Hankz Hankui Zhuo, Zhanwen Zhou

Keywords: genetic algorithm, reinforcement learning, deep learning, electronic paper

TL;DR: his paper proposes a novel GA-RL hybrid framework that efficiently optimizes E-paper driving waveforms under severe evaluation constraints by combining global evolutionary search with simulation-guided local fine-tuning.

Abstract: The calibration of driving waveform parameters is essential for achieving accurate color performance in electronic paper (E-paper) manufacturing. However, this process constitutes a high-dimensional black-box optimization problem, severely constrained by the fact that each real-hardware evaluation requires several minutes. This stringent evaluation budget renders manual tuning inefficient and limits the direct application of intelligent optimization algorithms. While Genetic Algorithms (GA) and Reinforcement Learning (RL) are principled approaches for such black-box optimization, GA can suffer from premature convergence and RL from sparse feedback under extremely limited evaluations. To address these challenges, we propose a novel GA-RL hybrid optimization framework. Our method embeds an RL-based local fine-tuning process, guided by a predictive LAB simulation model, within the global evolutionary cycle of GA. This cooperative integration enables efficient use of the limited evaluation budget. Experimental results demonstrate that the proposed framework significantly outperforms standalone GA or RL, achieving superior optimization efficiency and success rates in practical E-paper production.

Submission Number: 71