Defending Against Sophisticated Poisoning Attacks with RL-based Aggregation in Federated Learning