Fall 2013 Reinforcement Learning for P2P Backup Applications Shikhar Mall

A five year study of file-system medadata shows that the number of files increases by 200 % and only a select few file-types contribute for over 35 % of the files that exist on a file-system. 1 It is difficult to point out a permanent selection of files that a user really cares about. This project uses reinforcement learning (RL 12) to exploit the correlation between file-types, their usage patterns, multiple revisions etc., to extract out a selection of files which are “important ” for an individual user. In this project, we integrate this file-selection approach with an open-source P2P backup application called CommunityBackup. With this approach, such a backup application can auto-configure its file selection without the user having to update the selection every time s/he relocates his/her folders. The survey also points out that most file-systems are only half-full on average, independent of the user job category. 2 A P2