LABELPROP

Contacts: Kevin Montes

Short Description

Accelerate construction of physics event databases using a semi-supervised machine learning-based detector

Keywords

event, database, detection, machine-learning, semi-supervised, MDS+, TOKSEARCH

Long Description

The LABELPROP module is designed to accelerate the construction of large databases of physics events by using a semi-supervised machine learning algorithm together with manual analysis and validation. It relies on the SCOPE module to guide the user through generating a dataset of relevant signals for many shots. Shots in the dataset can then be manually analyzed using the included plot interface, with which the user can label the times at which the event of interest occurs. With few initially labeled examples, a semi-supervised learning algorithm searches through the entire dataset to find other shots and times at which the dynamics of the relevant signals resemble those characterizing the manually labeled events. In this way, information is ‘propagated’ from the few initial examples labeled by the user to potentially many unseen event instances. These detections of new events can then be manually validated by the user, and the process can be iterated repeatedly to build an event database up from scratch.

Typical workflow

The module workflow has three main steps:

Dataset Assembly

• Selection of relevant signals using SCOPE module

• Import from SCOPE in machine learning-friendly format

• Option to use example datasets to reproduce results from associated publication

Manual Analysis and Labeling

• Plot relevant signals for individual discharges with customized plotting options

• Prediction output for any already implemented detectors shows user where to look for event

• Quickly view events and add new ones with cursor via interactive plot features

Event Detection

• Choose shots to include as initially labeled examples and select event to detect

• Run algorithm with arbitrary choice of input parameters

• Run multiple algorithms scanning multiple combinations of input parameters

• Display detector’s performance statistics and correlate performance with the run/experiment type

Tutorials

• Google docs

• PDF

Classes

• omfit_classes.omfit_rdb.OMFITrdb()

Contributors

List of contributors sorted by number of lines authored:

2667 Kevin Montes
  32 Sterling Smith
   6 Orso Meneghini

Submodules

None

Users

List of usernames sorted by number of module imports: montesk, huwen, maris, marisa, ssmith, reac, stewarts, abbatej, anguloa, churchillr, mcclenaghanj, montes, sinclairg, sweeneyrm