An adaptive inference framework for coherent beam combining using deep reinforcement learning

Coherent beam combining (CBC) technology increases the peak power of laser systems by combining multiple coherent beams, overcoming the power limitations of single fiber lasers. This technology is essential for high-power laser applications in military and medical fields. Precise phase control is crucial for efficient beam combining, and reinforcement learning (RL) methods are being explored to address this challenge. This paper introduces a deep RL inference framework that supports various agents and includes an adaptive retraining mechanism to maintain performance amid changing environmental conditions. When a 90% drop in performance is detected, the framework triggers retraining using the improved proximal policy optimization (PPO) algorithm until performance recovers to 95%. Simulations show the framework can adapt to new environments within 1-3 training cycles, ensuring continuous learning and stable performance in dynamic settings.