Abstract-- We present and evaluate a reinforcement learningbased RWA algorithm for all-optical networks subject to physical impairments. The technique is suitable for decentralized networks and is compared with other techniques with similar computational complexity.