Post your week 14 discussion questions here as comments.
Week 14 Discussion Questions
6 Responses to Week 14 Discussion Questions
Leave a Reply
You must be logged in to post a comment.
Departments of Geography & Civil and Environ Engineering, UC Berkeley
From actors to agents in socio-ecological systems models
M. D. A. Rounsevell*, D. T. Robinson and D. Murray-Rust
Phil Trans Royal Soc B 2012
The authors open by stating: “The term ‘empirically grounded’ model is used here to refer to the systematic identification of process representations through inductive or deductive methods as well as calibration and validation of models that encapsulate these process representations.” In this view, empirical behavior ( captured via the imperfect instrument of social surveys) provides boundary conditions for human action. However, survey data is, by its nature, correlational, and much current “mixed methods” theory in the social sciences argues that “thick” methods are needed to understand causation. But causation is the “process” that the authors are trying to infer from survey data, and then extrapolate through modeling. What tare the implications of this assumption for both the internal and external validity of any findings?
Bounded rationality is an assumption of many (most?) ABMs, but is often insufficient to describe why people make particular decisions. Factors such as social status, religious or spiritual belief, or calculation of seemingly unrelated benefits may be more important. In what kinds of cases is the assumption of bounded rationality reasonable?
Agent typologies assume that each agent takes on only one role (e.g. farmer, politician, environmentalist), but many people wear multiple “hats” in decision making processes. Can ABM account for this phenomena of real systems?
Overall question: ABMs neglect community-level decisionmaking processes, politics, and much of the nuance of social systems. Even more than with environmental systems, it is crucial to ask whether the models bear any resemblance to the phenomena they are built to explain. Discuss!
Population growth and collapse in a multiagent model of the Kayenta Anasazi in Long House Valley
Robert L. Axtella,b, Joshua M. Epsteina,c, Jeffrey S. Deand,e,f, George J. Gumermane,f, Alan C. Swedlundg, Jason Harburgera,h, Shubha Chakravartya, Ross Hammonda,i, Jon Parkera,h, and Miles Parkera,j
This ABM relies entirely on demographic data (age at birth, age at death, age at marriage) that Roundsvell caution against. Does this reflect a belief that earlier societies are less complex than current ones?
What are some key differences between using ABMs to understand past responses to change and predict current societies’ attempts to cope with change?
Is contemporary society, with its skepticism about global limits to resource extraction and growth, qualitatively different from earlier societies such as the Anasazi, who exploited resources in a bounded geographic area? If so, what are the implications for ABM?
What ecological mechanisms that we have discussed in class might explain the model’s response to increased spatial (landscape) heterogeneity?
Agent-based modeling of deforestation in southern
Yucata ́n, Mexico, and reforestation in
the Midwest United States
Steven M. Manson*† and Tom Evans‡
The authors combine surveys data to identify and typify agents with experimental methods derived from game theory to characterize a decision-making space. Drawing on critiques voiced in class last week, how do the simplifying assumptions of game theoretics complicate or inform the picture that survey evidence provides?
The authors find that diversity in decisions is necessary to explain patterns of land use change in both study regions. How can this insight be used in a policy setting?
WEEK 14 (4/23/13): Agents in complex environmental systems
Rounsevell, M. D. A., D. T. Robinson, and D. Murray-Rust. 2012. From actors to agents in socio-ecological systems models. Phil. Trans. R. Soc. Lond. B. 367:259-269.
1. Rounsevell et al. propose the notion of human functional types (HFTs), as an analogy of plant functional types, to support the expansion (scaling) of ABM to larger areas. Do you think HFTs can account for the diversity of human behavior in addition to describing general functions?
2. How can ABM applied to human behavior be rigorous when qualitative methods based on coding approaches are so important, especially with respect to non-economic behavioral factors? Is this a case of subjectivity applied to modeling being an “art” as much as a science?
3. Which of the 3 scaling methods (scaling up, scaling out, and nesting) is most likely to best represent complex systems?
Manson, S. M., and T. Evans (2007), Agent-based modeling of deforestation in southern Yucatan, Mexico, and reforestation in the Midwest United States, Proceedings of the National Academy of Sciences, 104(52), 20678-20683.
1. Indiana households that reside on parcels ranging from 1 or 2 hectares to hundreds of hectares were studied. Aren’t the owners of small versus large parcels different because some might be investors? Does it make sense to call all owners “households”?
2. A key finding of this research is that models that focus solely on biophysical factors, such as topography or soil fertility, underemphasize the importance of social factors in local-level patterns of land change. How can the important social factors to consider be determined?
3. Wouldn’t the subjects chosen for the lab experiment influence the outcomes? Were they students? Or drawn from the real population?
4. The Yucatan models differed in two respects. HELIA agents preferred, in aggregate, to site agriculture on upland forest, whereas the econometric model found a weak negative relationship. Agents also uniformly ignored slope, whereas the econometric model found a positive relationship between slope and the probability of deforestation. Could these findings be used to suggest a flaw in a model? Which model is wrong? How can we determine when a model is not informative?
5. The authors say their household-level approach captures complexity and heterogeneity that is lost at higher levels of aggregation. Is aggregation always a bad thing? Do higher levels of aggregation tell us the wrong story?
Axtell, R. L., J. M. Epstein, J. S. Dean, G. J. Gumerman, A. C. Swedlund, J. Harburger, S. Chakravarty, R. Hammond, J. Parker, and M. Parker (2002), Population growth and collapse in a multiagent model of the Kayenta Anasazi in Long House Valley, Proceedings of the National Academy of Sciences of the United States of America, 99(Suppl 3), 7275-7279.
1. Why, as the authors suggest, is agent heterogeneity difficult to model mathematically?
2. When is qualitative agreement of model results with real data enough to substantiate a working hypothesis versus not sufficient? These authors found a model of the history of demographic changes and settlement patterns in Long House Valley to closely reproduce qualitative features but the model yielded populations that were substantially too large. Does this suggest a quantitative check should always be in place for these types of studies?
Axtell et al.
How many variables are enough to develop an heterogeneous model? How many calibrations (and assumptions) in your model do you need to do in order to follow an historical trajectory? Are the models representatives of the reality or are they follow the historical trajectories?
Manson & Evans
“In the Mexico case, we found that models of bounded rationality and perfect rationality produce similar results in aggregate, whereas the former can also disaggregate the rules of thumb used by individuals”… “In addition to representing bounded rationality, evolutionary programming allows agents in aggregate to replicate some characteristics of statistical models of perfect rationality while also individually deriving strategies that are typically identified through household interviews and qualitative research”. Which is the threshold between the aggregate and dissagregate in the scalar integration of the model?
“A key finding of the research is that models that focus solely on biophysical factors, such as topography or soil fertility, underemphasize the importance of social factors in local-level patterns of land change….Household attributes that are more difficult to measure with standard survey approaches, such as learning, information/ knowledge, risk aversion, and social networks, are hypothesized to play an important role in the heterogeneity of land management decisions.”
If the importance of social factor is key, how can this model represent the heterogeneity of landscape managed decision based in surveys that have a lack of information? Could this model represent a real scenario?
Rounsevell et al
How many attributes are necessary to develop a “typical agent”?
How any information do you lose in the scaling process? How can you deal with the correct scale of a model to predict the process?
1. How does one decide the unit size in ABM?
2. What implications do ABM efforts regarding reforestation (such as the results of Manson and Evans) have on reforestation policy? E.g., can Brazil work to establish jobs in other industries to avoid the rainforest logging that is threatening its indigenous tribes?
3. Manson and Evans mention evolution and survival of the fittest as being key factors in ABM; how do the intergenerational differences compare to intragenerational ones?
4. How well is it possible to correlate major disasters/large shifts in climate with archeological data which is used as input in ABM? E.g., can we tell when major droughts hit accurately enough to predict a large shift in settlement properties (clarify the use of the June Palmer Drought Severity Index)?
5. Is it difficult to utilize ABM in non-isolated environments? E.g., were the Kayenta Anasazi isolated enough from intertribal warfare or trade such that outsiders had little effect on their settlements?
6. What do ABMs look like (both in terms of simulation code and modeling output)? I somewhat envision them as complex versions of the game of life.
Rounsevell et al.
This is a fairly recent paper. What is the position of ABM in current efforts (especially in the realm of CHNS)? Are there objections to it not represented here?
What are the advantages/disadvantages of inductive vs. deductive approaches in the context of ABMs? What are the differences/functionality/utility of quantitative vs. qualitative methods?
How is identification of relevant agent characteristics accomplished?
How might PCA be used in conjunction with ABM to determine relevant characteristics? what insights from our readings on modeling and uncertainty can be applied to ABM?
“…heuristic strategies provide a transparent view of the decision-making process and therefore aid in the understanding of the human-environment relationships being modelled.”
Does this aid in the understanding of the system or rather make explicit the employed understanding of the system?
How would you choose which decision-making strategy to incorporate into an ABM?
“The difficulty in using AI algorithms to represent agent decision-making occurs when mapping the formulated algorithms or rules back to real-world behaviour or processes.” Is this difficulty mitigated if learning is done on real-world data? How might this difficulty be reduced?
Creating agent typologies with human actors (especially deductively) seems like it might run into sticky issues (think phrenology). Do you see this as a potential drawback of ABM? How might it be responsibly dealt with?
How could ABMs account for dynamic human types? (preferences, function, and strategy can change)
How would you go about validating a nested ABM such that it would be an effective/reliable tool for policy makers?
Axtell et al.
As with our readings on modeling and uncertainty, ABMs raise questions of equifinality. Are there differences/additional considerations with ABMs in this regard?
What are ways of quantitatively analyzing the differences in observed and modelled spatial distribution? Is it possible to do so in a way that can inform a further iteration of the model?
What do we learn from a model of a system where we know the real-world answers already?
What is the significance of the different Lp norms?
This is one way of looking at the usefulness of comparing models to real-world data: attribution of real-world behavior to factors outside of the model (i.e. social factors, in this case). How do you determine that the difference between model output and real-world behavior is due to factors not included in the model as opposed to poor model performance? How do you determine what unaccounted-for factors are responsible? What are the differences between the approaches of quantifying external factors through analyzing differences between model and real-world behavior and attempting to incorporate external factors into the model?
Manson & Evans
Increased heterogeneity was important to increasing model fit in both Manson & Evans and Axtell et al. Does increased heterogeneity increase specificity of a model at the cost of transferability? Can heterogeneity be thought of as a parameter in itself?
Both this model and Axtell et al.’s model seem like they underperform in terms of incorporating social factors. Is it possible to realistically incorporate social factors into an ABM? What are the challenges? How might they be addressed?
Is it possible/desirable to incorporate lab experiment results into the ABM? Or is this more of a demonstration/confirmation of the importance of diversity and suboptimal utility-maximization for understanding the system?
Generally speaking, Manson & Evans argue that there is something to be learned from many different methods, but how do the approaches inform each other?
How could this type of investigation of correlations through survey data and land-cover (or other geographic) data contribute to social science? What are the potential roadblocks?
What approaches are possible to deal with attributes not easily measured by surveys?
What is to be learned from model intercomparison (evolutionary and econometric)? Is mapping to real-world actors/data/dynamics necessary?
Manson & Evans cite “the importance of household factors in landscape outcomes and the potential drawbacks of methodological approaches that aggregate these local processes.”
Is this then an argument for a “nesting” approach as described by Rounsevell et al.? What alternatives are there?
Manson & Evans suggest that diversity of actor characteristics requires diversity of policies to effect change. Do you agree? How does this follow from their study? How can these types of models be used to inform policy creation?
Rounsevell et al 2012.
1. The use of typologies in ABM reminds me of advertising’s use of “big data” from social networking and Internet use to target marketing messages according to some cohort definition. Is there potential to use this data mining approach to ABM, as opposed to thinking of it as an extension of the survey?
2. Although they talk about the role of feedbacks, the outlined structure of ABMs seems to lead to a deterministic outcome (even with typologies) which seems not only boring and uninspiring, but also unrealistic in that it leaves out possibilities for revolution and innovation in a society. Is that a fair interpretation? The genetic programming and machine learning may include stochastic elements but is that a valid way to characterize social learning? How to overcome the “black box” problem so that social learning can occur but its mechanisms are also revealed? How to model a visit by the lone ranger, an external character/force that changes everything?
3. How would you model a revolution? The uprisings in Egypt, for example.
4. Representation of institutional and governance structures implies a hierarchy in agent-based modeling. Is it fair to represent these structures as parameters? In reality, interaction of government and agents seems messy, unpredictable, and very sensitive to local conditions, attitudes and politics. Can ABM help break down centralized governance approaches and lead us to new governance structures?
Manson and Evans
1. Perfect rationality versus bounded rationality: “While bounded rationality is a key form of decision making for individuals, we can also usefully make assumptions that fall under the aegis of perfect rationality (p. 20679).” So they both can lead to similar results, does that make them both useful modeling devices for different reasons or is equifinality an issue?
2. Which land use theories are more valid in what modeling situations: “relative space” vs. “absolute space?”
3. Learning land suitability patterns seems like a huge issue in reality and modeling. How to model? Do we need a better understanding of how this process works in order to model it? Is utility maximization an appropriate default approach?
4. How to overcome issues with aggregation without creating an overly complex model?
Axtell et al.
1. Fascinating, but seems oversimplified after reading Manson and Rounsevell. Is this another example of over-reliance on parameter tuning in order to meet desired results?
2. In Fig 3, the simulation results look very similar except for the outlying settlements. This seems like a common theme in modeling results. How important are these outliers?