Feature request: Kalman-inspired proposals (DREAM-KZS)

[wshlavacek]

Summary

Implement the Kalman-inspired proposal mechanism from DREAM(KZS) as described in Zhang, Vrugt et al. (2020), Water Resources Research, 56, e2019WR025474.

Background

DREAM(KZS) uses a Kalman gain matrix computed from the cross-covariance between parameter values and model outputs in the archive to steer proposals toward high-probability regions during burn-in. After a configurable fraction of burn-in (default 30%), the Kalman proposal is switched off and probabilities are renormalized between parallel direction and snooker updates.

Implementation considerations

This requires storing model outputs (not just parameters and scores) in the archive, and computing the cross-covariance matrix C(Z, f(Z)) and output covariance C(f(Z), f(Z)). Currently the algorithm only sees a scalar score per evaluation, not the raw simulation output vector. The data flow from the scheduler would need to be extended.

References

• Zhang, J., Vrugt, J.A., Shi, X., Lin, G., Wu, L. & Zeng, L. (2020). Improving Simulation Efficiency of MCMC for Inverse Modeling of Hydrologic Systems With a Kalman-Inspired Proposal Distribution. Water Resources Research, 56, e2019WR025474.