2018 October 30

Location: GRLA 107

Group members discussed Hayden’s project, which involves making decisions when provided a large number of earth models from a stochastic inversion. An outline of conversation content is below. Based on the meeting, we need to provide a tiny bit more structure to the one-slide conversation starters to bring focus towards the ML aspects of projects.

Attending: Andy Thomas Iga Jihyun Arnab Hayden Jonah

TODO:

  • Decide on a different meeting time using a Doodle poll (Jihyun)
  • Add Naive Bayes to Resources tab (Hayden)
  • Email website version of slide as a pdf to Thomas (Hayden)
  • Provide slide content guidelines to focus on ML aspects in meetups (Thomas)

Making decisions using a suite of stochastically inverted earth models (Hayden)

Goal: “When provided many earth models that all fit data, how do we make a decision?”
Input: Rock properties: density, p-velocity, s-velocity, and porosity
Framework: Naive Bayes Binary Classifier
Output: High vs. low producer classification
Details:
- Stochastic inversion input: rock physics model + amplitude-vs-offset
- Initial labels derived from user-defined production-based cutoff
- Many statistical assumptions in stochastic inversion framework
Problems:
- Consistent misclassifcation near faults.
- Dimensionality reduction problems.
Questions:
- How does Naive Bayes work?
- What are the statistical assumptions in the inversion? Are they violated by Naive Bayes?
Deltas:
- Consider exploring which of the 5,000 models predicts best?
- Consider using other classification algorithms RF, SVM, …
- Consider using smaller number of rock properties as inputs to lower dimensionality

Next week

Topics
- Bane - semester project
- Arnab - existing project
- Backup - Iga

Comments