230 likes | 407 Views
OS4316-06. A Hierarchical Process Model of the Ross Sea Ecosystem. Stuart R. Borrett, Will Bridewell and Pat Langley. Center for the Study of Language and Information, Stanford University. Kevin Arrigo. Department of Geophysics, Stanford University. February 2006
E N D
OS4316-06 A Hierarchical Process Model of the Ross Sea Ecosystem Stuart R. Borrett, Will Bridewell and Pat Langley Center for the Study of Language and Information, Stanford University Kevin Arrigo Department of Geophysics, Stanford University February 2006 Ocean Sciences Meeting, Honolulu, HI
Ross Sea Phytoplankton http://www.washington.edu/newsroom/news/images/antarctic-map.jpg Diatoms Phaeocystis antarctica What processes determinespecies dominance? Photoinhibition vs. Iron
Task Search for Models that Explain the Data Two Spaces • Structures • Beam search • Parameters • Gradient decent Inductive Process Modeling Given 1) Data (time-series) 2) BackgroundKnowledge • Entities • variables • parameters • Processes • hierarchical • functional forms • parameters Langley et al. in press; Asgharbeygi et al. in press; Todorovski et al. 2005
Processes Growth Predation Death Lotka-Volterra Library of Generic Processes
Ross Sea Data Forcing Observed SeaWiFS 1996-1997 JGOFS 1996-1997 CIAO Model (Worthen & Arrigo 2004, Arrigo et al. 2003, Tagliabue & Arrigo 2005)
Background Knowledge: Ross Sea Process Hierarchy Space of Possible Structures
Background Knowledge: Parameters Parameter Search Space
Two Trials • Trial 1 • Generic Processes Hierarchy • Parameters fairly constrained (CIAO) • Entities Phytoplankton, Zooplankton, Detritus, NO3, Fe, Env. • Trial 2 • No Fe
Trial 1 – with Fe Trial 2 – no Fe Simulations SSE = 205,520 r2 = 0.84 SSE = 1.36e+06 r2 = 0.70
Trial 2 SSE =1.36e+06r2 = 0.70 Trial 1b SSE =226,403r2 = 0.90 Trial 1 Trial 1a SSE =203,520r2 = 0.84
Trial 2 – no Fe Trial 1 – with Fe Model Fits Alternative structures may have similar fit Models with Fe are generally better
Future Work • Hierarchical entities • Spatial models • New criteria for selecting good models • Use additional system knowledge • Alternative parameter estimation algorithms • Process sensitivity
Summary & Conclusions • Modeling as search • H-IPM algorithm for automatic search • Reusable libraries of generic processes • Discover process models that explain the data • Multiple model structures may have similar fits • Models with Fe were better • Parameter sensitivity • Require more data to constrain the search
Acknowledgements • Arrigo Lab (Stanford) • Computational Learning Lab (Stanford) • NSF grant # IIS-0326059 sborrett@stanford.edu http://cll.stanford.edu/~sborrett/
Left Overs generic process Predation relates: R1{prey}, R2{predator} parameters: a[0,1], b[0,1] equations: d[R1,t,1] = a * R1 * R2 d[R2,t,1] = b * R1 * R2
Phytoplankton in the Ross Sea SeaWiFS, Sept. 1997-Aug. 1998 Diatoms Phaeocystis antarctica What processes determinespecies dominance? Photoinhibition vs. Iron