This set of code runs experiments over real world networks. Selenium on top of a PyVirtualDisplay is used and the actual graphics is disabled.

Trained RL model needs to be saved in rl_server/results/. We provided a sample pretrained model with linear QoE as the reward signal. It can be loaded by setting NN_MODEL = '../rl_server/results/pretrain_linear_reward.ckpt' in rl_server/rl_server_no_training.py.

RL, robustMPC, MPC are implemented in rl_servers/. Other ABR schemes, namely BB, RB, Festival, BOLA and DASH original, are natively supported in dash.js/, where the switch abrAlgo can be found in dash.js/src/streaming/controllers/AbrController.js. These algorithms are called from specific HTML files in video_server/. Experiments run over RL, robustMPC, MPC and BOLA in random shuffles.

To conduct the experiment, modify url in run_video.py to the server address, and then run

python run_exp.py

To view the results, modify SCHEMES in plot_results.py (it checks the file name of the log and matches to the corresponding ABR algorithm), then run

python plot_results.py