EXPLORING THE LINKS BETWEEN DECISION MAKING AND ENVIRONMENTAL DEGRADATION USING AGENT-BASED MODELING: THE PREHISTORIC ARGRICULTURAL LANDUSE MODEL

Matthew Peeples, Robert Cox, and Tanjot Bhatia


TABLE OF CONTENTS



 INTRODUCTION

A great deal of recent research in a variety of disciplines has been directed towards understanding how humans use resources and solve social dilemmas in the context of dynamic socio-ecological systems. Increasingly, agent based and systems dynamics models have played a major role in this research (e.g., Anderies 2000; Parker et al. 2003; van der Leeuw 2004; Castella et al. 2005). These formal methods for modeling human-environmental interaction allow for the exploration of the internal and external factors effecting stability and change in complex socio-ecological systems which cannot be observed or studied directly. In this way, simulation provides a means of identifying the key variables and key relationships between variables that can provide a great deal of information regarding human behavior.

Archaeologists interested in dynamic socio-ecological systems in the prehistoric past can greatly benefit from the application of simulation studies. The laboratory of the long and diverse archaeological record is uniquely suited to examining the human use of resources specifically because the time-scale covered is immensely greater than that observable through other methods. Changes in the prehistoric past are, however, primarily observable in their consequences at the macro-level. Thus, modeling provides a means of exploring the possible role of micro-level processes on producing these macro-level dynamics (van der Leeuw 2004).


THE PROBLEM

One of the fundamental questions addressed by archaeologists studying prehistoric landuse is how the decisions of small social units in agricultural societies potentially lead to large scale environmental degradation. Many past explorations of human impacts on their physical environments have focused on environmental variability and population density as the major drivers of degradation (e.g., Kohler and Mathews 1988; Zubrow 1971). It is true that under most circumstances, a larger population has a greater potential to outstrip its natural resource base. This explanation does not, however, address an essential and closely related question. Why do people remain or move into a degrading environment despite resource depletion and increasing labor costs, even when other options are available (see Hegmon and Kintigh 2005)? Previous attempts to address this question have drawn on theoretical concepts relating to the role of risk and uncertainty in human decision making and the “sunk costs” derived from physical (or psychological) investments in a particular location (Kohler 1992; Rautman 1993; Varien 1999; Janssen et al. 2003). In the context of this paper, we attempt to investigate the potential role of a related concept, the social attrativeness of an environment, in anthropogenic environmental degradation (Hegmon and Kintigh 2005).

Farming, especially in small scale societies, is inherently a social activity. Thus, decisions regarding where to move or when to remain in a degrading environment are fundamentally influenced by the perceived need to maintain social connections to a place and to other people. Not all places on the landscape are created equally, however (environmentally or socially). Over time, certain locations develop into more popular places due to both population and their lengthy histories. Thus, they can become more attractive from a social standpoint, even when they are ecologically or economically less attractive. This is what Hegmon and Kintigh (2005) have referred to as the “Manhattan Effect.” We will explore the potential effects of this issue, among others, through the development of an agent based model of prehistoric agricultural landuse.


THE PREHISTORIC AGRICULTURAL LANDUSE MODEL

The Prehistoric Agricultural Landuse Model (PALM) presented here is an effort to create a simple, expandable, and configurable agent based model of agricultural landuse in the Pueblo region of the prehistoric American Southwest. This model was developed using the NetLogo platform and is based on the Sugarscape model presented by Epstein and Axtell in their influential book Growing Artificial Societies (1996). PALM is also similar, in many respects, to the Artificial Anasazi [1] (Dean et al. 2000; Axtell et al. 2002; Gumerman et al. 2003) and Village Project (Van West 1994; Kohler et al. 2000; Reynolds et al. 2004) models which have been developed to examine landuse in specific portions of the prehistoric Colorado Plateau. The approach we have taken in the conceptualization of this model, however, is more general. We have developed a small number of simple variables through which various assumptions regarding decision making and environmental variability can be explored in different environments. Thus, this model is not directed towards replicating reality or even directly replicating observed macro-level data. Rather, this model is an attempt to better understand the potential relationships between human decisions and natural environments as well as to identify new variables which may not have been previously considered in archaeological studies.

Based on previous anthropological, agronomic, and archaeological examinations of small scale agricultural societies throughout the world (Netting 1993; Kohler and Matthews 1988; Kohler 1992; Sandor et al. 1990; Van West 1994; Fish 2004), we have developed a simple conceptual model of agricultural landuse which we have implemented into PALM (Figure 1). As Figure 1 shows, we have positively related resource degradation to both the total population and the length of time agents spend in a particular settlement. Importantly, the social attractiveness of a settlement is also positively related to these same two variables. Thus, we see this relationship as having both positive and negative feedbacks. Beyond this, resource degradation leads to increasing hardships (declining yields, increasing labor, etc.). The flow of population between settlements in our conceptual model is essentially related to changes in the level of resource degradation, the population, and the social attractiveness of every settlement. Thus, through this model we will be able to examine the potential role of mobility in the degradation (or lack thereof) of the resource base. The decision models of agents in our model are implemented in relation to this common conceptual background.

Implementation

The landscape of the PALM model is designed so that various resource levels and configurations can easily be explored in relation to different assumptions regarding the actions of agents. Currently, we have developed four landscapes, two of which will be discussed in the context of this paper (Figure 2). Each landscape is represented by series of discrete clusters of individual patch resources, representing agricultural land. Each of these discrete clusters is a “settlement” or agricultural catchment which is essentially a completely separate landscape. The distance between settlements on the map is not meant to represent actual geographic distance, but the settlements are conceptualized as distant enough that agents cannot freely move between them.

The patches within each settlement have a productivity value representing the patch’s vulnerability to degradation as well as its productive potential. Resources are degraded based on the intensity of use at each time step an agent remains in a particular patch. Intensity is calculated randomly within a range which can be varied. Thus, the productive potential of a patch declines with both the intensity and length of use. The relationship between intensity, length of use, and declining yields is supported by a great deal of research into small scale agriculture throughout the world (Netting 1993; Kohler and Matthews 1988; Kohler 1992). Resources in patches regenerate logistically when they are not being used, but if they are completely degraded, they will no longer regenerate.

Each settlement also has a variable representing the social attractiveness, which is equal to the sum of the total time all agents currently in the settlement have spent in this location. In this way, as suggested above, social attractiveness relates to both the popularity of the place (i.e. total population) and its history (i.e. length of occupation for each agent). We have chosen not to model the specifics of labor investments or individual benefits of social attractiveness in this version of the model, but this may be implemented in the future.

Agents in the model represent single households, which are the smallest units usually visible in the archaeological record (see also Gumerman et al. 2003). In the current version, population is held constant. All agents remain in the model for all time steps and do not “die” even when they do not always meet their resource requirements. This was done for the sake of simplicity and in order to make the effects of other variables easier to interpret. Each household has a size randomly assigned to one of three values. This is then used to assign a resource requirement for households of each size. Households start out randomly assigned to a settlement. Agents can see the patches within a radius of 10 patches around themselves. This was the largest vision radius that could be used due to computational constraints.

At each time step, households move to a random one of the closest empty patches within their vision radius which has the highest productive potential. If their current patch has the highest productive potential, they do not move. Households then harvest resources, consume resources, and store surplus resources. The stored resources a household has can be declined at a variable rate to represent food which rots in storage, is eaten by insects, lost, or saved for planting (see Van West 1994:123-126). The amount of resources harvested at each time step is related to the productive potential of the patch the household inhabits as well as a random value within a range which can be varied. Thus, the total harvest and storage varies greatly from patch to patch and from time step to time step. This is a reasonable approximation of the arid Southwest, where areas with favorable precipitation and temperatures are extremely localized and vary from year to year (Fish 2004:118-119).

After a number of time steps, defined randomly for each agent within a range, agents reevaluate their current situation and decide whether to change settlements or stay in their current settlement. [2] In addition to this, if the current settlement of an agent doesn’t contain at least one patch with resources for every time step until they can consider migrating again, the agent will move. Agents evaluate their situation using information relating to their current settlement and all other settlements. In the current version of the model, agents have perfect information on all settlements, but this will probably be changed in the future (see conclusions).

Agents make their decisions to stay or move in order to maximize either the resources per household or the total social attractiveness of their settlement. Resources per household are used rather than total resources so that settlements can be compared without regard to total area. This is essentially the average resources available for each agent within the settlement. Agents also consider their current storage in their resource level. Thus, stored resources can provide a buffer to resource declines and migration. When an agent migrates to a new settlement, all of their stored resources are lost. Thus, migration is a costly option. Agents considering social attractiveness instead of resources, simply choose to move if another settlement has a higher social attractiveness value than their current settlement, regardless of resource considerations. The probability of choosing to maximize either social attractiveness or resources can be varied using the social attractiveness decision parameter.

Each time step in the model is meant to represent 1 year. The model is set to run for a total of 200 time steps or until all resources are depleted. Thus, we are able to track changes in the results of agent decisions over time. All of the variables described here, as well as additional variables, are explained in more detail in Table 1.

Testing the Model

Initial testing of PALM consisted of informally varying parameters to ensure that the variables all worked correctly in relation to one another. This informal testing revealed that there was little variation in the results between runs with the same initial settings. Additionally, this revealed that a limited number of variables were responsible for a majority of the meaningful variation observed in the behavior of the model. Thus, due to computational and time constraints, we chose to limit formal testing to 5 runs per configuration. We also decided that it was not feasible to do systematic sweeps of every variable incorporated into the model. We therefore concentrated on the most important variables as determined by the informal experimentation.

Formal testing was carried out by making incremental changes to individual variables for both landscapes 1 and 2, with all other variables held constant at their default values (see Table 1). Table 2 shows all configurations which were tested. As not all of these tests provided useful and interpretable results, only a few of the tests will be discussed in the context of this paper. Since we are concerned with issues relating to both mobility and environmental degradation, we chose to use the resource level and the average number of migrations per year as the major indicators of model behavior.


RESULTS OF THE MODEL

The testing procedure described above revealed a number of interesting relationships between the variables considered and the behavior of the model. In this section, we will discuss the results which are the most pertinent to the current study. We conclude this paper by providing future directions for further analysis and development of PALM.

As discussed above, we are interested in identifying the effects of micro-level decisions on the part of individual households on the macro-level behaviors in the model. We proposed above that the decisions of agents regarding social attractiveness, as we have described it here, may be one possible causal factor in degradation of the resource base of a particular settlement. In order to provide a baseline of comparison, however, it is first important to summarize the behavior of the model without social attractiveness included.

In our initial discussion of the problem of resource use in agricultural societies, we hypothesized that, while total population probably has important effects for the degradation of agricultural resources, this alone is not adequate to describe how and why degradation takes place. It is thus important to first examine the effects of the total population level on the behavior of the model. Expectedly, as the results of testing show, there is a positive relationship between resource degradation and total population in the PALM model (Figure 3). This, in and of itself, however, does not really provide any useful interpretive insights into how the decisions of agents can lead to degradation. Thus, other aspects need to be considered.

Informal testing of the model suggested that the maximum intensity of landuse is an important causal factor in resource degradation. As Figure 4 shows, when the maximum possible intensity increases, the resource base of both landscapes experience increasing degradation. Importantly, however, within the range of values tested, none of the settlements were completely degraded. This suggests that, even when landuse is extremely intense , agents acting only on information regarding the number of agents per resource level in each settlement are able to distribute themselves across the landscapes in a way that does not lead to severe degradation. [3]

In addition to this, testing indicates that the maximum possible harvest has important effects on the behavior of the model. As Figure 5 shows, however, increasing the maximum possible harvest of the model environment does not drastically effect the degradation of the resources of either landscape. On the other hand, the maximum harvest is strongly related to the average number of migrations per year. Figure 6 shows that, as the maximum harvest increases, the number of migrations per year declines rapidly. This is likely due to the fact that, in a more productive environment, agents are more frequently able to build up a buffer of stored resources which allows them to stay in place. In addition to this, Figure 7 shows that there is a major threshold point in the maximum productive potential values, past which the number of migrations per year decline rapidly. Interestingly, this threshold is approximately 3.0. This value is twice the mean resource requirement for each household. Thus, this suggests that when agents are able to store resources beyond their own requirements more than about 50% of the time, migration declines greatly. Intriguingly, ethnographers studying the modern Pueblo communities note that households often have the goal of producing about twice their annual resource requirements every year (Titiev 1944:181; Bradfield 1971:121; Forde 1931:393).

As the paragraphs above suggest, when all agents choose their location based purely on resource level per household and the total population is set at the default level, no scenario was produced where the resource base of any settlement was completely degraded. When agents are allowed to choose their location based on the social attractiveness of each settlement, however, the behavior of the model changes significantly. Figure 8 and Figure 9 demonstrate that, as the probability of choosing a location based on its social attractiveness increases, the level of resource degradation increases in both landscapes. In landscape 2, where all settlements are close in total resource level, each settlement appears to be affected about equally. In landscape 1, however, where settlement 3 is much larger than all others, settlement 3 becomes severely degraded more frequently at lower values for the social attractiveness decision parameter. This is likely due to the fact that, as settlement 3 is able to sustain a larger population for longer than the other settlements, the social attractiveness of this place (sum of the lengths of occupation for all current agents) continually increases, leading to a positive feedback where agents deciding where to move based on social attractiveness are more likely to move to this settlement.

Social attractiveness is also strongly related to the average number of migrations per year. As Figure 10 shows, when the probability of choosing a location based on social attractiveness increases, the average number of migrations per year increases rapidly. In addition to this, the number of migrations increases rapidly as resources are depleted. In landscape 2, this appears to take place at a much lower level for the social attractiveness decision parameter. Figure 11 clarifies this point. In landscape 1, the average number of migrations increases steadily until agents are choosing their location based on social attractiveness about 85% of the time. At this point the migrations sky rocket and the resource base is depleted. In landscape 2, on the other hand, migrations increase sharply when agents are choosing their locations based on social attractiveness as few as 30% of the time. This suggests that the distribution of resources on the landscape can drastically affect the thresholds at which social attractiveness becomes a major factor in resource degradation.


DISCUSSION AND CONCLUSIONS

As this discussion has illustrated, population level alone is not an adequate explanation for the degradation of resources in agricultural societies. In this model, drastically different degrees of resource depletion were simulated with the same stable population level simply by varying the decision parameters of individual agents. In the modeling environment, when information is perfect and resources are the only thing agents take into consideration, they were able to distribute themselves evenly across the model landscape and sustain themselves for 200 simulated years without severely damaging any specific areas of their landscapes.

On the other hand, when agents are given the option of deciding where to move based on the social attractiveness (the popularity and history) of a particular settlement, severe degradation can occur rapidly and across the entire landscape. Importantly, the results here show that differences in the amount and distribution of agricultural resources within the simulated landscapes can lead to major differences in the threshold at which rapid degradation occurs. Additionally, when resource levels are small, only a small proportion of decisions on where to move need to be based on social attractiveness for severe degradation to occur. PALM has also demonstrated that, even in an extremely variable (spatially and temporally) landscape, if the maximum potential harvest is higher, the average storage increases and agents are more likely to stay in one place for a longer period of time. All of these results suggest interesting relationships between the environmental characteristics of landscapes, the decisions of individual social units, and resource depletion which could be further examined by expanding the PALM modeling framework.


FUTURE DIRECTIONS: RE-IMPLEMENTATION OF PALM

Since the completion of the initial model, PALM has been re-implemented in the RePast platform. This change allows for overall speed increases in running a PALM simulation, as well as providing an expandable framework for incorporating new factors that affect agents’ decisions in a dynamic landscape.

The current development version of PALM implements several features that allow for further exploration of environmental and decision parameters. These new features include:

1) The ability to easily create varying resource concentrations, ranging from large to small resource areas and from dense to dispersed resources
2) Calibration of resource patch productivity values to Southwestern soil data
3) Simulation of several long-term resource depletion scenarios
4) Simulation of various rainfall scenarios based upon Southwestern paleoclimatic data
5) Testing of different strategies agents employ in deciding when and if to migrate to another settlement.
6) Experimental imperfect information networks that limit an agent’s choices for migration
7) Population growth

These new features expand the capabilities of the model, and the implementation of the new features allows for a more detailed examination of the interaction between a dynamic environment and the changing decision parameters of agents that adapt to local knowledge of resources. Furthermore, the new implementation of PALM is further grounded in archaeological data, such as with the incorporation of long-term soil depletion into the model


REFERENCES CITED

Anderies, J. M. 2000. On modeling human behavior and institutions in simple ecological economic systems. Ecological Economics 35:393-312.

Axtell, R., J. Dean, J. Epstein, G. J. Gumerman, A. Swedlund, S. Chakravarty, and M. Parker. 2002. Population Growth and Collapse in a Multi-Agent Model of the Kayenta Anasazi in Long House Valley. Proceedings of the National Academy of Sciences 99(3):7275-7279.

Bradfield, M. 1971. The Changing Pattern of Hopi Agriculture. Royal Anthropological Institute Occasional Paper 30. London.

Castella, J., T. N. Trung, and S. Boissau. 2005. Participatory Simulation of Land-Use Changes in the Northern Mountains of Vietnam: the Combined Use of an Agent-Based Model, a Role-Playing Game, and a Geographic Information System. Ecology and Society 10(1).

Dean, J., G. J. Gumerman, J. M. Epstein, R. L. Axtell, A. C. Swedlund, M. T. Parker, and S. McCarroll. 2000. Understanding Anasazi Culture Change through Agent-Based Modeling. In Dynamics in Human and Primate Societies: Agent-Based Modeling of Social and Spatial Processes, edited by G. J. Gumerman and T. A. Kohler, pp. 179-205. Santa Fe Institute and Oxford University Press, Oxford.

Epstein, J. M. and R. L. Axtell. 1996. Growing artificial societies: Social science from the bottom up. MIT press and the Brookings Institution, Cambridge.

Fish, S. K. 2004. Corn, crops, and cultivation in the Greater Southwest. In People and Plants of Ancient Western North America, edited by Paul Minnis, pp. 115-166. Smithsonian Institution Press, Washington , D.C

Forde, C. D. 1931. Hopi Agriculture and Land Ownership. Journal of the Royal Anthropological Institute 61:357-406.

Gumerman, G. J. 2003. The Evolution of Social Behavior in the Prehistoric American Southwest. Artificial Life 9:435-444.

Hegmon, M. and K. W. Kintigh. 2004. Why Did They Stay? Diversity in Decision-Making Regarding Residential Mobility. Paper presented at the 104th annual meeting of the American Anthropological Association, Washington, D.C.

Janssen, M. A., T. A. Kohler, and M. Scheffer. 2003. Sunk-Cost Effects and Vulnerability to Collapse in Ancient Societies. Current Anthropology 44(5):722-728.

Kohler, T. A. 1992. Field Houses, Villages, and the Tragedy of the Commons in the Early Northern
Anasazi Southwest. American Antiquity 57(4):617-634.

Kohler, T. A. and M. H. Matthews. 1988. Long-Term Anasazi Land Use and Forest Reduction - A Case Study from Southwest Colorado. American Antiquity 53(3):537-564.

Kohler, T. A., J. Kresl, C. Van West, E. Carr, and R. Wilshusen. 2000. Be There Then: A Modeling Approach to Settlement Determinant and Spatial Efficiency among late Ancestral Pueblo Populations of the Mesa Verde Region, U.S. Southwest. In Dynamics in Human and Primate Societies: Agent-Based Modeling of Social and Spatial Processes, edited by G. J. Gumerman and T. A. Kohler, pp. 145-178. Santa Fe Institute and Oxford University Press, Oxford.

Netting, R. M. 1993. Smallholders, Householders: Farm Families and the Ecology of Intensive, Sustainable Agriculture. Stanford University Press, Stanford.

Parker, D. C., S. M. Manson, M. A. Janssen, M. J. Hoffman, and P. Deadman. 2003. Multi-agent Systems for the Simulation of Land-use and Land-cover Change: A Review. Annals of the Association of American Geographers 93(2):314-337.

Rautman, A. 1993. Resource Variability, Risk, and the Structure of Social Networks: An Example from the Prehistoric Southwest. American Antiquity 58(3):403-424.

Reynolds, R. G., T. A. Kohler, and Z. Kobti. 2004. The Effects of Generalized Reciprocal Exchange on the Resilience of Social Networks: An Example from the Prehispanic Mesa Verde Region. Computational & Mathematical Organization Theory 9:227-254.

Sandor, J. A., P. L. Gersper and J. W. Hawley. 1990. Prehistoric Agricultural Terraces and Soils in the Mimbres Area, New Mexico. World Archaeology 22(1):70-86.

Titiev, M. 1944. Old Oraibi, A Study of the Hopi Indians of Third Mesa. Papers of the Peabody Museum of Archaeology and Ethnology Vol. 22, No. 1. Cambridge, Massachusetts.

Van der Leeuw, S. E. 2004. Why Model? Cybernetics and Systems: An International Journal 35:117-128.

Van West, C. R. 1994. Modeling Prehistoric Agricultural Productivity in Southwestern Colorado: A GIS Approach. Reports of Investigations 67. Department of Anthropology, Washington State University, Pullman and Crow Canyon Archaeological Center, Cortez, Colorado.

Varien, M. D. 1999. Sedentism and Mobility in a Social Landscape. University of Arizona Press, Tucson.

Zubrow, E. B. W. 1971. Carrying capacity and dynamic equilibrium in the Prehistoric Southwest. American Antiquity 36:127-138.


END NOTES

1. The Artificial Anasazi model was also originally developed based on Sugarscape.

2. We implemented migration in this way primarily due to computational constraints. Allowing agents to choose migration at any point slowed the program significantly.

3. In the tests run here, the highest intensity value used allows agents to degrade a patch by as much as 50% in a single time-step.



Stat Counter