Structural realism and theory development in agent-based models addressing practical problems Volker Grimm
Steve Railsback ACKNOWLEDGEMENTS Department of Ecological Modelling (ÖSA) at UFZ, Leipzig Teachers and inspirators: Christian Wissel, Janusz Uchmanski, Don DeAngelis Collaborators and students
What do I mean by “theory”? “A scientific theory is a well-substantiated explanation of some aspect of the natural world that is acquired through the scientific method and repeatedly tested and confirmed through observation and experimentation. … As used in everyday non-scientific speech, "theory" implies that something is an unsubstantiated and speculative guess, conjecture, or hypothesis; such a usage is the opposite of a scientific theory.” Wikipedia 11.5.2016 Agent-based complex systems science
THEORY IS GOOD FOR APPLICATION
Overview of this lecture Structural realism Pattern-oriented modelling Pattern-oriented theory development First principles Individual-based ecology
Idea underlying all modelling Real world/system too complex to understand or to predict behaviour Create a simplified representation, which only contains essentials with regard to a certain question or problem Understand and predict behaviour of this simplified representation Transfer this understanding and these predictions to the real system
Problem: verification and validation the model “mimics the real world well enough for its stated purpose (Giere, 1991)” (Rykiel 1996, p. 230). V2 we can place confidence “in inferences about the real system that are based on model results (Curry et al., 1989)” (Rykiel 1996, p. 230) Note: Rykiel combines both aspects under one term, validation Rykiel 1996
Hildenbrandt et al. 2010
GENERATIVE MECHANISMS We want to make sure that our models are “sufficiently good” representations of their real counterparts. We want to learn about the real world We want to capture essential elements of a real system’s “internal organization” We want to capture the “generative mechanisms” that generate the structure and behaviour of real systems
Predictive ecology Only if we capture the “generative mechanisms” sufficiently well will our predictions be good enough for new conditions Bossel (1992) contrasts descriptive models with “real-structure” models: “The difference is that between the […] descriptive modelling of the motion of the hands of a clock, and the analysis and real-structure description of the mechanism of the clock; only the latter would be able to predict correctly what would happen if the pendulum were stopped or if the spring were not rewound.” (p. 264).
Predictive systems models should be STRUCTURALLY REALISTIC STRUCTURAL REALISM Predictive systems models should be STRUCTURALLY REALISTIC Reproduce observed patterns for the right reasons, i.e., capture the internal organisation (across scales) of the real system Test: Independent predictions! Optimize model complexity („sweet spot“, „middle ground“)
Models should reproduce patterns, not data General relationships that preferably hold across different instances of the same system Robust relationships Structures or processes that characterize a certain class of systems Related concept in economics: „stylized facts“
Spatial patterns in ecology http://www.gov.nf.ca/nfmuseum/images/empetrumnigrumlivedeadwaveforestmistakenpoint.jpg
Spatial patterns in marine ecology Tremblay et al. (unpublished)
More patterns …
What scientists do with patterns Pattern: Beyond random variation Patterns contain information about internal organization We develop models that reproduce the pattern We infer from the mechanism built into the model the real system´s organization We need to exploit („squeeze“) the pattern
Problem with complex systems A single pattern may not contain enough information Ecologists tend to focus on (single) patterns observed at one level of observation Behaviour Population dynamics Community composition Ecosystem function
“Monoscopic” view Most approaches (and modellers) are not making the best use of the information (lemons) available
The thing we need is a “multiscope”
Multiscope view Take into account multiple patterns Observed at different scales and/or levels of organisation Make your model reproduce these patterns simultaneously Use each pattern as a „filter“ to reject unacceptable submodels or parameterizations „Pattern-oriented modelling“(Grimm et al. 1996, 2005; Wiegand et al. 2003, 2004; Grimm and Railsback 2005, Railsback and Grimm 2012).
Pattern-oriented modelling
Example: Oystercatcher mortality (1976-1981) Pattern-oriented modelling (POM) Example: Oystercatcher mortality (1976-1981) Definition: „Multi-criteria design, assessment, and parameterization of models of complex systems“
Patterns as filters Multiple (3 or more) „weak“ patterns may narrow down model structure better than one single „strong“ pattern Cycles in small mammals („strong“) Abundance within certain bounds Recovery after disturbance needs 10 years Territory size changes with abundance in a certain way
Creative scientists (Sherlock Holmes) are using POM all the time Patterns: Examples Red shift in spectra of galaxies and stars Atomic spectra Iridium layer: mass extinctions DNA: Chargaff‘s rule, x-ray diffraction patterns, tautomeric properties of building blocks Periodic system of elements Creative scientists (Sherlock Holmes) are using POM all the time
Pattern-oriented Modelling: Three elements Design: Provide state variables (entities, processes) so that patterns observed in reality in principle also can emerge in the model Model selection: Use multiple patterns for contrasting alternative (sub)models of certain adaptive behaviours Parameterization: Use multiple patterns for determining entire sets of unknown parameters („inverse modelling“)
Pattern-oriened theory development Pattern-oriented Modelling: Three elements Design: Provide state variables (entities, processes) so that patterns observed in reality in principle also can emerge in the model Model selection: Use multiple patterns for contrasting alternative (sub)models of certain adaptive behaviours Parameterization: Use multiple patterns for determining entire sets of unknown parameters („inverse modelling“) Pattern-oriened theory development
Pattern-oriented theory development Theory in Individual-based Ecology is across-levels Theory=models of what individuals do that explain system dynamics (Capture enough essence of individual behavior to model the system)
THEORY DEVELOPMENT CYCLE Proposed theories: alternative traits for a key agent behavior ABM How well does ABM reproduce observed patterns? Empirical literature, research Characteristic patterns of emergent behavior
EXAMPLE: VULTURES AND CARCASSES Pattern: # of feeders at a carcass Jackson et al. 2008. Biology Letters 4 Cortes-Avizanda A, Jovani R, Donázar JA, Grimm V. Ecology (2014)
EXAMPLE: VULTURES AND CARCASSES Cortes-Avizanda, Jovani, Donázar & Grimm. 2014. Ecology.
EXAMPLE: VULTURES AND CARCASSES Cortes-Avizanda, Jovani, Donázar & Grimm. 2014. Ecology.
EXAMPLE: VULTURES AND CARCASSES Cortes-Avizanda, Jovani, Donázar & Grimm. 2014. Ecology.
Better for pattern-oriented theory development: FIRST PRINCIPLES Often, we base our theories on ad hoc assumptions. Better for pattern-oriented theory development: Start from “first principles” Physics, chemistry Fitness seeking
First principles: example Benjamin Martin (PhD student, UFZ) Daphnia population dynamics in laboratory Effects of pesticides My idea: Start from existing model Ben: That model is good, but everything is based on empirical (imposed, calibrated) relationships I want to to do something more generic I want to try Dynamic Budget Theory (DEB)
DEB – Kooijman 2010 where and
DEB Growth Food Maintenance Reproduction
DEB-IBM: Generic NetLogo program Martin et al. (2012) Methods Ecology and Evolution Population Environment Toxicants Food Temperature Individual DEB IBM Growth Reproduction Survival Density Size distribution DEB-IBM
Parameterization
We could reproduce not only population density at one food level, but density and size distribution for multiple food levels and toxicant expsoure Low food (0.5mgC d-1) High food (1.3mgC d-1) Martin et al. 2013. Am. Nat. Data from Preuss et al. 2009
ANOTHER INDEPENDENT PREDICTION Martin, Jager, Nisbet, Preuss & Grimm. 2013. Am. Nat.
Ecology (populations, communities, ecosystems) Emergent Specific environment Population growth rate, λ Vital rates, b & d Wide range of environmental conditions Adaptive behavior (IBM) Imposed Empirical Ecology (populations, communities, ecosystems)
Generic models of interaction: THEORIES OF WHAT Generic models of interaction: Zone-of-influence approach Forest gap models: vertical competition for light (JABOWA) Size-based trophic, „trait-mediated“ Generic models of behaviours Foraging Habitat selection Home range Generic models of life history Bioenergetic models Ontogenetic Growth Model Dynamic Energy Budget theory
EXAMPLE: TROUT HABITAT SELECTION
GENERAL THEORY Railsback and Harvey 2013. Trends Ecol. Evolution.
Pattern-oriented modelling Pattern-oriented theory development So far: Structural realism Pattern-oriented modelling Pattern-oriented theory development First principles Individual-based ecology
INDIVIDUAL-BASED ECOLOGY 2015. BioScience 65: 140-150
INDIVIDUAL-BASED ECOLOGY
Phase 1: Conceptualization
Phase 2: Implementation
Phase 3: Diversification
Coherence for IBE
SUMMARY Role of theory for application Structural realism: generative mechanisms Pattern-oriented modelling: multiple patterns as filters Pattern-oriented theory development: models of behaviours that explain system-leven patterns First principles: evolution, physics, chemistry Individual-based ecology: ultimately big science