Input-dependent baseline for reducing the variance from external input processes.

Paper: https://openreview.net/forum?id=Hyg1G2AqtQ

• Regular A2C with state-dependent baseline on the load-balancing environment

python3 load_balance_actor_critic_train.py --num_workers 10 --service_rates 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95 1.05 --result_folder ./results/regular_value_network/ --model_folder ./results/parameters/regular_value_network/

• A2C with multi-value baseline (10 value networks) on the load-balancing environment

python3 load_balance_actor_multi_critic_train.py --num_workers 10 --service_rates 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95 1.05 --result_folder ./results/10_value_networks/ --model_folder ./results/parameters/10_value_networks/

• Monitor learning progress: Tensorboard in ./results/, policy perforamnce on unseen traces plotted as test_performance.png in ./results/parameters/.

• Example results: training Tensorboard screenshot in ./figures/training.png, testing plots in ./figures/regular_value_network_testing.png and ./figures/10_value_networks_testing.png.

Python 3.6, Tensorflow 1.2.1, Numpy 1.14.5