Side-channel attacks exploit secondary information, such as power consumption, to extract sensitive cryptographic keys. Although machine learning (ML) methods have enhanced these attacks, their reliance on extensive, device-specific training data limits cross-device applicability and assumes unrealistic levels of target system access. In this work, we propose a novel domain adaptation strategy based on Procrustes Analysis to enable robust ML-based side-channel attacks across intra-model devices. Complementing this approach, we introduce a reinforcement learning framework to elevate dataset entropy, thus significantly reducing the training data volume required. Evaluation on ten TI Tiva C microcontrollers executing the Data Encryption Standard (DES) and eight Atmel XMEGA 128A1U microcontrollers implementing the Advanced Encryption Standard (AES) demonstrates significant improvements in model accuracy and a substantial reduction in guessing entropy. Notably, our RL-generated datasets achieved a higher average training dataset entropy than traditional methods, providing a richer model training environment.