#CS176 HMM Project

Background

This is a parallized implementation of parameter estimation of a hidden markov model using the Baum-Welch algorithm. There is a special harness mrca.py to utilize the parameter estimation for estimating the TMRCA problem.

Running MRCA

You can run EM on a new set of initial parameters and a FASTA file with the following:

python mrca.py initial_params.txt fasta_sequence.txt output_estimated_params output_initial_decodings output_estimated_decodings output_likelihood

This will run 15 iterations of EM and output the corresponding files. If you wish to run more iterations or specify a threshold, see the call to tmrca.estimate_params(max_iter=15) in mrca.py. You can specify either a threshold for the change in log-likelihood or a maximum iteration (or both!).

Example:
tmrca.estimate_params(thresh=1e-3, max_iter=100)

Code Overview

class TSequence

TSequence is a wrapper abstraction for performing initial decodings, parameter estimation, and estimated decodings for the TMRCA problem. Every TSequence is associated with an initial parameter and observed difference sequence. The class will automatically compute the observed differences.

TSequence.estimate_params

This function must be called before calling decode_estimate and/or accessing the the following class attributes: estimate, likelihood, and initial_likelihood.

class EM

Generically performs parameter estimation for a given initial parameter and observed sequence. The code has been parallelized to calculate transition and emission probabilities.

class Decoding

Wrapper for calculating all decodings: Viterbi, posterior decoding, and mean expectation.

plot module

Utilizes Matplotlib to plot the estimated decodings for the mu, 2mu, 4mu, and 5mu samples. Call plot_initial or plot_estimate with mu, 2mu, 4mu, or 5mu to see the plots.

Example
$ python -i plot.py
>>> plot_estimate('mu')

NOTE: You must have matplotlib installed in order to plot.

Solutions to PS5 (estimates)

Our estimations and plots are stored in the /estimates/ folder.

##Authors

• Anthony Sutardja
• Kevin Tee