Wednesday, March 14, 2018

Grey's view on grand challenge.

As for a ‘grand challenge’, I see the biggest open challenge as about how to build models that learn. That is, how do we leverage the power of machine learning and modern inference techniques for learning multi-scale physical and emergent principles in watersheds and complex systems? How do we construct models that allow for direct learning form data, but also allow us to prescribe what we do know about the biogeophysics of ecohydrological systems?

1 comment:

  1. Agreed. A group of us think the basis for that lies in an approach that draws upon Probabilistic (where probabilities represent beliefs) and Information Theoretic philosophies and concepts. To learn, model must have the ability to morph/evolve its structure (not just its parameters and state variables). Information in the form of evidence/data (whatever the form of the evidence) must be transformed into information about conceptual variables (fluxes, states, boundary conditions, parameters etc.) and eventually into information in the form of models (which are relationships between conceptual variables). Finally models are used to transform information in the form of evidence about one kind of conceptual variable (drivers/inputs) into information (often predictive) about other kinds of conceptual variables (state variables/outputs). In all of this, information contained in evidence is always about relationships. The challenge is to figure out how to represent those relationships in ways that can morph with the evidence.