Scaling RL Training 20x: Lessons from Distributed Systems at Ericsson

Designing high-throughput distributed RL architectures with GPU learners, 100+ CPU actors, and multi-node coordination for radio network optimization.