@apvalentine.bsky.social
Assoc. Prof., Durham University Inverse theory, ML, seismology, geophysics & computation. https://valentineap.github.io/
An ensemble resampler provides an efficient means of constructing a trans-conceptual ensemble from multiple single-concept ensembles.
In particular, if we have already run McMC for each state individually, Carlin & Chib's approach lets us build a trans-conceptual ensemble by resampling the resulting ensembles.
Go read doi.org/10.1029/2024... to find out more and see some examples of this idea applied to real problems!
(4/4)
Carlin & Chib (1995) combine separate states into a single 'product space' and add an indicator variable to identify the active state.
An alternative comes from the work of Carlin & Chib (1999). Instead of working across multiple different parameter spaces, combine them all into a single 'product space'. Of course, this increases the dimension of the search space (= bad), but in practice some simplification may be possible.
(3/4)
Reversible jump McMC can explore, and step between, parameter spaces of different dimension.
In some cases this problem can be tackled using reversible-jump McMC (Green, 1995) to switch between the different physical theories (or 'conceptual states'). See, e.g., 'trans-dimensional' sampling. However, this requires us to know the Jacobian of the transformation between states...
(2/4)
An illustration of three classes of ensemble representing an Earth model whose properties only vary with depth. The fixed-dimensional ensemble shows three Earth models each with three layers, whose thicknesses and seismic velocities all vary. The trans-dimensional ensemble is similar but with variable numbers of layers. In contrast, the trans-conceptual ensemble allows Earth models to have variable character as well. The first containing a single linear gradient, the second multiple layers and an inclined fault, while the third has undulating layers. Each ensemble class is a subset of those to its right. Fig 1 from Sambridge, Valentine & Hauser, 2025, JGR.
🧪 Monte Carlo methods can help characterise unknown parameters within a physical model. But what if we have multiple competing physical theories? Trans-conceptual sampling may be the answer...
Just out in JGR, w/M. Sambridge & J. Hauser: doi.org/10.1029/2024...
@durhamearthsci.bsky.social
(1/4)
Very flattered by this invitation to deliver the 2025
@royalastrosoc.bsky.social
Harold Jeffreys lecture - in honour of a
Durham University alumnus who made seminal contributions to geophysics and probability theory, and who also happens to be my great-great-grand-PhD-advisor! 🧪
@durhamearthsci.bsky.social staff and students build the world. Expect to see the result at open days and beyond!
04.11.2024 09:52 — 👍 4 🔁 1 💬 0 📌 0
👋 Can I be added to the science feed, please? This is me: scholar.google.com/citations?us...
10.10.2024 16:15 — 👍 2 🔁 0 💬 2 📌 0
TWO!! preprints out today from our work on #Bayesian inference:
1. Matthias Scheiter's #PhD research: #AI-enhanced McMC - essopenarchive.org/users/841077...
2. New paradigm: 'Trans-conceptual sampling' w/
M.Sambridge & J.Hauser - essopenarchive.org/users/841079...
@durhamearthsci.bsky.social
