Implemented convergence tests and output the thinned chains to file (instead of doing a production run). The first burn-in (max 3000 steps) allows the walkers to settle. Once the chains have settled, I move on to the actual burn-in run. The walkers are "settled" once they reach a threshold for delta(logP) and delta(position). I noticed that sometimes after the first burn-in, walkers will split off into local minima. So, after the first burn-in, I reinitialize the walkers to the position of the "best walker" and re-start the burn-in from there (initialized randomly within 1% of the best position). Then, I use the auto-correlation time to determine when the walkers have converged. The chains are thinned by their auto-correlation time and written to file.
Implemented convergence tests and output the thinned chains to file (instead of doing a production run). The first burn-in (max 3000 steps) allows the walkers to settle. Once the chains have settled, I move on to the actual burn-in run. The walkers are "settled" once they reach a threshold for delta(logP) and delta(position). I noticed that sometimes after the first burn-in, walkers will split off into local minima. So, after the first burn-in, I reinitialize the walkers to the position of the "best walker" and re-start the burn-in from there (initialized randomly within 1% of the best position). Then, I use the auto-correlation time to determine when the walkers have converged. The chains are thinned by their auto-correlation time and written to file.