vishwakftw / Pump-Failure-Prediction

Coursework pertaining to CS5350 : Bayesian Data Analysis offered in Spring 2019

Pump Failure Prediction

What is the goal of the problem?

• You have 10 power plants pumps. For these 10 pumps, you are given the number of failures (x_i) and the number of hours in thousands that the pumps have run for (t_i).

• We seek to model the relationship between x_i and t_i using a Gamma-Poisson Hierarchical model as done by George et al in their paper Conjugate Likelihood Distributions in 1993.

What is the goal of the assignment?

• We will be solving this problem using Metropolis-Hastings algorithm - a popular variant of the Monte-Carlo Markov Chain (MCMC) algorithm, which will be implemented natively using "standard" libraries.

• Our implementation will then be compared against an implementation in PyMC3, a popular probabilistic programming framework.

• The sampling algorithm will be compared against an variational inference algorithm which is optimization based. This has been implemented using PyMC3 as well.

Further details

• Further details of the problem and comparison of implementations are presented in the report.

Running the code

• Please install the requirements specified in requirements.txt. To use the PyMC3 implementations, you will need to install PyMC3.
• python <script> --help will print a set of configurable options while running the code where <script> is one of [numpy_mcmc.py, pymc3_mcmc.py, pymc3_vi.py].
• The dataset for the assignment is here

References

• OpenBUGS description: http://www.openbugs.net/Examples/Pumps.html
• George E.I., Makov U.E., Smith A.F.M., Conjugate Likelihood Distributions, Scandinavian Journal of Statistics, 1993.

---

Done as part of coursework pertaining to CS5350 : Bayesian Data Analysis offered in Spring 2019