deb-kit2 / online-sparse-approximation

Greedy Algorithms for Online Sparse Approximation

Run for :

• HTP
• CoSaMP
• Iterative Hard Thresholding
• Subspace Pursuit : noisy

Inplemented for :

• HTP
• CoSaMP
• Subspace Pursuit
• Iterative Hard Thresholding

To-Do :

• Refine readMe inside logs folder.
• Re-run regret experiment with corrected IHT.
• Be insightful about the variances of all random things.
• If doesn't converge, try with tau = 100 fixed.
• plots : reward vs time
• plots : regret vs time
• plots : average regret vs time
• plots : accumulated reward vs time
• make notebook.
• Other algorithms.
• Fix x_best and continue programming.

Observations :

• Converges with the log range of tau.

Unanswered :

• Should I take np.abs() in the algorithms?

Questions :

• In regret calculation step, is x_best a constant through time?
• y_t should be varying, not fixed. Try both.
• Is gamma fixed?

Answers

• No. It changes with time. Obviously.
• y_t will vary as x_best varies. I was wrong about the understanding. Now the setting asks me to try with w_random as fixed and varying.
• Sir said, fixing it is valid for agnostic adversary scenario.