A 16-page report containing explanations of our most promising strategies, and results.
Our final deliverable presentation, outlining the content of the report.
See this file for an example of how to use our simulator to reproduce results.
This is where the simulator engine is located. The methods are detailed by the docstrings in simulator.py.
This directory contains output from experiments, in addition to the associated test code. See final_results.txt
for agreggated simulator output.
Contains code for the weighted running average strategy. See gametheory.py
for strategy classes.
Relevant Files:
clever_brute_force.py
brute force solverlinear_heuristic.py
linear heuristic approximation methodlinear_programming.py
old linear programming approach (does not work)
Relevant Files:
oneshot.py
contains the oneshot algorithmmultishot.py
contains the multishot algorithm
Relevant Files:
randomForestSimulator.py
Simulator filetensorflowVersion/Random Forest Tensorflow.ipynb
Research notebookrandomForest/
Scrapped haskell version, never tested
Contains code for the Vowpal Wabbit strategy.
Relevant Files:
formatting_script.py
formats our raw data into files with trainable examples for VW.plot.ipynb
contains graphs of data used in finding the best multipliers and models.simulate.py
contains theVWSimulator
class, which simulates revenue gained by using a VW model.testing.py
is used to tune for the optimal multiplier values and models.tuning.py
was used to train VW models on bids between two times.- The
models
folder contains the final models we used, with 1-5 passes through data.
Includes optimized classes in runnning_average.py
for experimenting with global and separated running averages.
Contains code that formats the data from s3://adsnative-sigmoid
Relevant Files:
filterer.py
filters out direct auctions and redundant bid linescombiner.py
reduces the number of files so as to reduce GET requests to S3