Keywords
NEAT, meta-modeling, ensemble neaural networks, uncertainty quantification
Start Date
26-6-2018 2:00 PM
End Date
26-6-2018 3:20 PM
Abstract
Applications of physically-based environmental models originating from research should be ubiquitous to use in both research and planning/consulting environments. However, due to their complexity, data resolution requirements, parameter number, platform affinity, and other criteria they are rarely suited “out-of the box” for field applications. Results from physically-based models are considered the most accurate but operating an entire system requires dedicated knowledge, extensive set up, and sometimes significant computational time. Questions from field applications conversely require easy to get, quick and “accurate enough” results. The use of web-services might alleviate some of the implications for model users but ultimately shift the responsibility and workload to the hosting environment.
To help closing the gap between research and field models we propose a machine learning (ML)-based meta-model approach aiming to capture the intrinsic knowledge of a physical model into an ensemble system of artificial neural networks and make it available for providing simplified answers to on the field problem-specific questions. A meta modeling approach was developed to help transitioning from research to field by enabling a modeling framework to interact with ML libraries to emerge model surrogates a(ny) modelling solution. The Cloud Services Integration Platform CSIP/OMS was extended and utilized to harvest data and derive the meta-model. Here, NeuroEvolution of Augmenting Topology (NEAT) techniques in an ensemble application, combined with ANN uncertainty quantification are the main methodologies used. Two examples applications have been prototyped and will be presented, a sheet and rill erosion model and a daily runoff model.
Framework-enabled Meta-Modeling
Applications of physically-based environmental models originating from research should be ubiquitous to use in both research and planning/consulting environments. However, due to their complexity, data resolution requirements, parameter number, platform affinity, and other criteria they are rarely suited “out-of the box” for field applications. Results from physically-based models are considered the most accurate but operating an entire system requires dedicated knowledge, extensive set up, and sometimes significant computational time. Questions from field applications conversely require easy to get, quick and “accurate enough” results. The use of web-services might alleviate some of the implications for model users but ultimately shift the responsibility and workload to the hosting environment.
To help closing the gap between research and field models we propose a machine learning (ML)-based meta-model approach aiming to capture the intrinsic knowledge of a physical model into an ensemble system of artificial neural networks and make it available for providing simplified answers to on the field problem-specific questions. A meta modeling approach was developed to help transitioning from research to field by enabling a modeling framework to interact with ML libraries to emerge model surrogates a(ny) modelling solution. The Cloud Services Integration Platform CSIP/OMS was extended and utilized to harvest data and derive the meta-model. Here, NeuroEvolution of Augmenting Topology (NEAT) techniques in an ensemble application, combined with ANN uncertainty quantification are the main methodologies used. Two examples applications have been prototyped and will be presented, a sheet and rill erosion model and a daily runoff model.
Stream and Session
Stream A, Session A3