Deep Reinforcement Learning Approach for Lagrangian Control: Improving Freeway Bottleneck Throughput Via Variable Speed Limit