/Optimm

Primary LanguageMATLAB

Datasets

Old datasets: Tapback and Recall New datasets: Recollect 2016 and 2017

HMM

To train the HMM:

Run train_hmm('recollect') to train the HMM on recollect

Run train_hmm('tapback') to train the HMM on recall

The follwing list of functions are called:

  1. setup_newDS() loads the data for recollect
  2. setup_oldDS() loads the data for tapback
  3. compute_emission_prob() returns the emission matrix B
  4. baum_welch_cont() finds the transition matrix A and the initial probabilities Pi Parameters
  5. reestimate_A() provides MLE for A and Pi
  6. single_seq() computes expected sufficient statistics for each training sequence
  7. alpha_beta_pass() carries out the forward-backward procedure
  8. forward2() computes the prediction
  9. compute_rmse() computes the rmse and the predictions
  10. control_fn() computes the beta distribution to compute emission probabilities.

To test the HMM:

Run test_hmm(A,Pi,accu,nbacks,nTrials), Values returned are: Test_Predicted_Values, Test_Actual_Values, test_rmse, test_B

To obtain predicted n-level trajectory plots from HMM:

Run explot3(A,Pi,test_B,test_n_backs_list,Test_Predicted_Values,Test_Actual_Values). The plots are saved into 'User_Skill_Trace.pdf'

To compare models:

explot2() generates plots that correspond to the mean and the variance of the error residuals across all subjects

To tune global item-parameters (rho, c, d, n) for HMM:

Run tuning_parameters(). Does a grid-search. Adjust suitable ranges for each hyper-parameter if needed.

To train HMM version 2:

NOTE: Here, theta is 3D. The transition function is theta(X-X', X'-nb_prev, binned(prev_acc)).

Run train_hmm2('recollect') to train the HMM on recollect

The follwing list of functions are called:

  1. computeA_B() returns the emission matrix B and the list of transition matrices A for each subject
  2. baum_welch_cont2() finds the transition probabilities theta and the initial probabilities Pi Parameters
  3. Estep() provides MLE
  4. single_seq2() computes expected sufficient statistics for each training sequence
  5. populateA() populates the transition matrix A at each time-step from theta

To write matlab objects into csv files:

Run write_file()

To Cluster the sequences via EM on the HMM:

  1. Run get_cluster_centers(): Current number of clusters = 3, can be changed with more data
  2. pre_train() : Initializing the cluster centers
  3. baum_welch_cont_EM() : performs EM clustering
  4. reestimate_A_EM(): Re-estimates cluster centers
  5. assignCluster(): Assigns a sequence to a cluster

To analyze clusters:

Run analysis_subj16() and analysis_subj17() to analyze clusters formed in Recollect-2016 and Recollect-2017 respectively.

NOTE: The Tapback and Recall datasets test subjects that go upto a n-back of 9. The memory score is 0.25, hence state space is 36, with 3 intermediate states between each level.

UKF model

To train the UKF model:

Run UKFTrain('recollect') to train the UKF model on recollect

Run UKFTrain('tapback') to train theUKF model on recall

The follwing list of functions are called:

  1. setup_newDS() loads the data for recollect
  2. setup_oldDS() loads the data for tapback
  3. learn_kalman() finds the parameters: A, B, Q, R ,X0, V0 via EM
  4. Estep() computes the expected sufficient statistics
  5. ukalman_filter() computes filtered estimates
  6. rts_smoother() computes smoothened estimates
  7. calc_sigmapoints() computes the sigma points for the unscented transform.
  8. g() is the non-linear function in the observation model

To test the UKF model:

Run compute_rmse(), that computes the rmse and the predictions

To obtain predicted n-level trajectory plots from the UKF model:

Run explot3(). The plots are saved into 'User_Skill_Trace.pdf'

To Cluster the sequences via EM on the UKF model:

NOTE: this doesn't converge. It can be tried with larger datasets.

Run UKF_EM() to cluster sequences. Current number of clusters = 2, can be changed with more data. It calls learn_kalman_EM.