Choosing parameters

• There are “setters” for some parameters, what are the most impactful knobs to tune?

  To increase accuracy, you might want to increase nm (defaults to 10) or the neighborhood size (defaults to 2) with set_nm and set_k_th_neighbor_to_search, respectively. The value of nm will affect the number of merges per block in the off-by-one \(e\)-matrix merging step. And the k_th_neighbor_to_search (call it k) parameter controls the number of points to check around a suspected local minimum. For example, setting k=2 means we have to do the (more demanding) graph partitioning steps for (2k+1) * (2k+1) = 25 grid points, with the suspected point lying at the center.

  To increase efficiency, you might want to increase c (defaults to 3) with set_c. This will enable the algorithm to skip multiple graph partitioning steps, at the cost of being prone to overshoot and getting trapped in a local optimum.

• How could I choose init_ka, init_kb, and i_th?

  You do not need to do that. These values are automatically determined by the algorithm. But they are helpful for debug. :-)