HcwXd / Discrete-HMM-Implementation

An implementation of discrete Hidden Markov Model

Discrete-HMM-Implementation

Execution

For compiling the two main program: train.cpp and test.cpp, simply run

make

Then we have two executable file train and test. For counting the accuracy of the predictoin from these two programs, you also have to run

g++ accuracy.cpp -o acc

Now, you have train, test , acc three files. These are all you need to execute the two shell script files HMM_processing.sh and Multiple_HMM_processing.sh. Also, you can simply run train and test to see the result respectively.

Train

./train ITERATION INPUT_INIT_MODEL INPUT_SEQ OUTPUT_MODEL

• ITERATION the integer indicates how many iterations the training will run
• INPUT_INIT_MODEL the file name of your initial model
• INPUT_SEQ the file name of your training data
• OUTPUT_MODEL the file name of your output model
• e.g.
  • ./train 30 model_init.txt seq_model_01.txt model_01.txt

Test

./test MODEL_LIST TEST_DATA RESULT

• MODEL_LIST the file name of the text file contains all the model you want to test
• TEST_DATA the file name of the data you want to test
• RESULT the file name of the output prediction by testing the data
• e.g.
  • ./test modellist.txt testing_data1.txt result.txt

Accuracy

./acc RESULT ANSWER

• RESULT the file name of the output prediction by testing the data
• ANSWERthe file name of the answer to the prediction by testing the data
• e.g.
  • ./acc result.txt testing_answer.txt

Running whole process of HMM

To run the whole process of HMM you can execute the two shell script files HMM_processing.sh and Multiple_HMM_processing.sh. The former one can run the whole process with certain number of iterations, the latter one can run several numbers of iterations at the same time to let you see the correlation between iterations and accuracy.

./HMM_processing.sh ITERATIONS 

• ITERATIONS the integer indicates how many iterations the training will run

./Multiple_HMM_processing.sh ITERATIONS1 ITERATIONS2 ITERATIONS3 ...

• ITERATIONS# the series of integers the training process will run

Result

From training and testing the data with different parameters, we can observe how the accuracy will change with different times of iteration.

The accuracy actually has a strike drop from 0.766 to 0.5364 with iteration equals to 1 and 10 respectively. Then the accuracy bounce back to 0.7852 when iteration equals to 20. After that, the accuracy grows quite steadily, and after iteration above 700, the accuracy become stable at 0.8692 with itearations equals to 1,500 and 2,000.

In brief, the data indicates that the maximum improvement we can make by adjusting iteration may happened around iteration equals to around 700 ~ 800. And with the increase in iteration above 800, the improvement in the accuracy become subtle.