Summary
Implement the multi-try variant of DREAM (MT-DREAM(ZS)) as described in Laloy & Vrugt (2012), Water Resources Research, 48, W01526.
Background
MT-DREAM generates k parallel proposals per chain, evaluates all of them, then selects one with probability proportional to the posterior density. The acceptance criterion uses a modified ratio that accounts for the multi-try selection.
This improves acceptance rates at the cost of k model evaluations per chain per iteration (instead of 1).
Implementation considerations
This is an architectural change to the scheduler interaction. Currently got_result processes one evaluation per chain per generation. MT-DREAM would require k evaluations per chain before making the accept/reject decision. The got_result / proposal generation flow would need significant rework to buffer multiple evaluations per chain.
References
- Laloy, E. & Vrugt, J.A. (2012). High-dimensional posterior exploration of hydrologic models using multiple-try DREAM(ZS) and high-performance computing. Water Resources Research, 48, W01526.
Summary
Implement the multi-try variant of DREAM (MT-DREAM(ZS)) as described in Laloy & Vrugt (2012), Water Resources Research, 48, W01526.
Background
MT-DREAM generates k parallel proposals per chain, evaluates all of them, then selects one with probability proportional to the posterior density. The acceptance criterion uses a modified ratio that accounts for the multi-try selection.
This improves acceptance rates at the cost of k model evaluations per chain per iteration (instead of 1).
Implementation considerations
This is an architectural change to the scheduler interaction. Currently
got_resultprocesses one evaluation per chain per generation. MT-DREAM would require k evaluations per chain before making the accept/reject decision. Thegot_result/ proposal generation flow would need significant rework to buffer multiple evaluations per chain.References