ranf / adx-bob Goto Github PK

From the specifications (page 15):

The way the game server simulates the daily real-time bidding on behalf of the
Ad Networks is as follows: Upon a bid request as a result of user u visiting web site
w, the matching Bid Bundle entries are set according to actual user attributes and the
user classification service level in effect²⁸.
²⁸Multiple segments may apply - resulting in more than one matching bid-bundle entry.

Implemented solution is assuming that if the campaign's market segment includes multiple segments than only the people who belong to both populations are targeted.

Need to investigate exactly how the segments are used and update the logic accordingly.

Store campaign reports and associate them with the parameters used to bid. Try to visualize relation between campaign profit, bid, and parameters.

Current code include unused classes and always returns a random bid.

BTW, please follow java naming conventions, Google Style Guide may be used for reference.
Consider adding unit tests as well.

Go over the TODOs (Window → Show View → Tasks in eclipse), fix the trivial ones and open issues for the rest.

http://www.javaslang.io/

Should provide easier, readable, testable approach for manipulating collection and java functional programming.
Need to use its classes instead of java.util.

Store previous bids (associated with campaign) and their results.
Use a distance function between current bid's campaign, and previous bids campaigns - determine if they are similar using a threshold.
Winning bids allow to us use lower bid, and losing bids require higher bid.
Apply some aggregation function to all similar campaigns bids to determine the new bid (make sure it's not too extreme).

Need to configure standard environment for the project.

1. Move eclipse project to root folder, use source code for adx-server instead of jar.
2. Configure build.gradle on root folder and on subprojects (some links).
3. Integrate with Travis CI.
4. Add code coverage metrics using Codecov.

Consider using kNN classification for bids.

• Separate bids to several classes.
• To find correct classification of past bids we need to save all parameter used to construct the bid and relevant game status, and when the game ends use the log parser to find what was the best possible winning bid (minimal for bundle / maximal for campaign).
• Build kNN model. Use various distance function between campaigns to find the best model.
• In-game we can use the model to find the winning bid.

Note that for this approach to work for campaign bids we also need to classify each campaign opportunity to profitable/not and achievable/not, and only try to win profitable+achievable campaigns.

• Same kNN approach should be much easier here - classification based on actual campaign results.

For each day, for each campaign, we need to store the base bid (without coefficients) and bid report (of next day). Then we will able to push them to the learn component.
Blocks #18

See improved revenue. Add tests for appropriate areas.

Write unit test.
Remove redundant / unreasonable code.

Current campaign bid performs pretty bad. Add tests and make sure the agent is able to regularly obtain at least 8 campaigns.

Add code to run bob against last year agents. Share results.

Receive many and worthy campaigns.

Different publishers have different market segment probabilities. When the UCS level is low it makes sense to bid more to these publishers.

Need to find out how queries to unknown market segment are calculated.