This repo consists python code for the OR paper "The best of many worlds: Dual mirror descent for online allocation problems".
nrm.py contains useful functions for run the experiments. time_period.py, resource_dimension.py, and decision_dimension.py are the code to run the experiments and generate output in csv files for the three subfigures in Figure 1, respectively. Here are the documents for the optional arguments in these three files:
optional arguments:
-h, --help: Show this help message and exit.
--T: The number of samples.
--num_trials: The number of random trials.
--num_params: The number of random parameters.
--step_size_constant: The step size is step_size_constant/sqrt{T}.
--reference: Reference function in mirror descent
--budget_ratio: The ratio between budget and the average consumption.
adx-alloc-data-2014 folder contains the data file generated following the procedure of "Yield Optimization of Display Advertising with Ad Exchange", Management Science. The dataset has 12 advertisers and 100,000 impressions. pub2-ads.txt contains the value of rho for each advertiser. pub2-sample.txt contains the revenue of matching each impression to the corresponding advertiser. We rescale the revenue so that the largest term is 1 in our experiment. We rescale \rho such that sum_j rho_j =1.5.
main.py is the main file to run the experiments, and save the output in csv files. Here are the documents for each optional arguments:
optional arguments:
-h, --help: Show this help message and exit. --num_trials: The number of random trials. --lambd: The coefficient of regularizer. --data_name: The name of the dataset. Must be pub1-pub7. --step_size_constant: The step-size constant. --regularizer: The regularizer r. --reference: The reference function h. --save_frequency: How many iterations we save for the output. --T_ending: The number of samples. --num_T: The number of T values. --sum_rho: The sum of rho.
cvxpy 1.0.31 pandas 0.20.2 numpy 1.16.6