FEARLUS: An Agent-Based Model of Land Use Change
The FEARLUS (Framework for the Evaluation and Assessment of Regional Land Use Scenarios) project started in April 1998, to run initially for five years. It is one of three main approaches to the study of land use change being conducted at the Macaulay Institute, the other two being analytical and empirical. The FEARLUS approach is to use agent-based simulation modelling, with the intention of eventually being able to provide advice to policymakers in the Scottish Executive on the possible effects of such things as regulation, climate change, and globalisation on land use change.
At the moment, FEARLUS is mainly involved in proof-of-concept work, trying to establish the advantages and disadvantages of agent-based modelling, and the kinds of problem it can be used to address. Thus, whilst FEARLUS is ostensibly focused on land use issues in Scotland (population ~5M, area ~8Mha) or particular subregions or catchments therein, at present the model itself is too abstract to be regarded as representing any particular area or time period.
The main reason for retaining as abstract as possible a model, which underpins the core methodology in the FEARLUS project, is the idea that explanations of emergent phenomena should be based on the simplest possible models. Two main approaches to agent-based modelling are emerging in the literature, one deriving from complex systems and cellular automata research, the other from distributed artificial intelligence — with the interaction between them focused on debate about realism versus tractability (Goldspink, 2000). Whilst the appropriate level of realism to use may in the end depend on the purpose of the model, the attitude within FEARLUS is that extra realism should only be built in on the basis of a thorough understanding of the dynamics of the interactions and the kinds of emergent phenomena that are generated by simpler models.
As it stands, the FEARLUS model is implemented in Objective-C using the Swarm simulation libraries. The agents are land managers, who have to decide each year which land uses they will use on the one or more land parcels they own. The decision is made using a decision algorithm, which is applied to each land parcel owned individually, rather than for the farm as a whole (if the land manager owns more than one land parcel, for example). The decision algorithm consists of three possibly different strategies, which reflect the context and behavioural aspects of the land manager. The first strategy is the contentment strategy. This strategy is used to determine the land use for a land parcel whose yield exceeds the land manager’s individual contentment threshold (for example, a habit strategy might be used in this case, which just applies the same land use as the previous year). If the yield is less than the contentment threshold, then the land manager has an individual propensity to either imitate or innovate. The decision algorithm therefore specifies an innovative and an imitative strategy for the land manager to use when the yield is unsatisfactory, together with a probability to determine which of these will be used each time.
Imitative strategies use exclusively information from neighbouring land managers’ land parcels and the land managers themselves when determining land uses — meaning that the set of land uses available for selection consists only of those that appear in the neighbourhood. For the purposes of imitation, a distinction is made between the social and physical neighbourhoods. The physical neighbourhood reflects the topological layout of the environment — which land parcels neighbour which other land parcels. Land managers, however, are simulated as exchanging information socially — thus the information drawn on when using an imitative strategy includes all land parcels owned by land managers with neighbouring land parcels to those owned by the land manager making the decision. An example of an imitative strategy might be to apply the land use that the majority of the land manager’s neighbours are using.
Innovative strategies make no use of neighbouring information, but may choose from any of the land uses. An example of an innovative strategy might be to choose a new land use at random — "innovative" for our purposes meaning that a land use may be introduced that is not currently being applied by any land manager in the neighbourhood.
Land managers may be grouped into sub-populations according to the decision algorithm they use. This is used to compare various decision algorithms for their competitive advantage in different environments. This competitive advantage is usually assessed on the basis of which sub-population owns the greatest number of land parcels at a predetermined time after the beginning of the simulation.
The environment consists of a uniform 2D grid of cells, each cell representing a land parcel — meaning land parcels all have the same area. Facilities are provided for simulating hexagonal and triangular cells, as well as squares with von Neumann or Moore neighbourhoods. The grid may be bounded, with edge and corner cells having fewer neighbours than the other cells, or toroidal ("wrap-around") in which edge cells have neighbours on the opposite edge. Each land parcel has individual biophysical properties, simulated using a string of binary digits ("bitstring" henceforth). These spatially varying biophysical properties remain constant during the course of the simulation — they are not affected by the land uses or climate. Temporal variation is introduced by the spatially homogenous climate and economy, also simulated using bitstrings. Since there is no difference between their function in the simulation, these may be referred to generically as "external conditions". A fixed set of land uses is determined at the start of the simulation, and all land uses are available for selection by land managers at all times (at least, by those using innovative strategies). Land uses are also simulated using bitstrings. The yield from a particular land use is determined by how well its bitstring matches with a concatenation of those of the external conditions and the biophysical properties of the land parcel.
Land managers accumulate wealth from the yield generated by their land parcels, less a constant break-even threshold, applied equally to the yield from all land parcels. Land managers with negative accumulated wealth must sell off their land parcels at a fixed, constant price, until their wealth is zero or above. If they lose all of their land parcels in this way, then they are removed from the simulation. Land parcels put up for sale are transferred to other land managers by choosing at random from the set of land managers with sufficient wealth owning neighbouring land parcels to the one that is for sale, and one new land manager. A land manager chosen to have the land parcel transferred to them will have their wealth deducted by the land parcel price. Land managers have no option to refuse this transfer.
The focus of work in FEARLUS so far has been on the competitive hierarchy of decision algorithms, and in particular, on purely imitative decision algorithms versus those with an innovative component (Polhill, Gotts, & Law, 2001). These studies have found that whilst the competitive advantage of various decision algorithms depends on the physical and social context, purely imitative strategies tend not to perform so well as those with an innovative component. A followup paper, currently under way, will look at aspiration level (another way of looking at the contentment threshold) and look for cases when imitation has an advantage over decision algorithms that make no use of imitation.
In the near future, the FEARLUS project will be starting work studying common-pool resource dilemmas, with a particular focus on the EU Water Framework Directive. This work will explore the use of multi-dimensional utility functions in land managers — separating the financial gains from using over-exploitative strategies from the social costs that might be applied by land managers who are affected in response. We also hope to introduce policy into the model, to see what kinds of regulation can be used to prevent over-exploitation of common-pool resources, and to look at possible warning signs by looking for the kinds of environment that promote over-exploitative strategies. A forthcoming review paper (currently being refereed) contains a survey of agent-based research in common-pool resource dilemmas (Gotts, Polhill, & Law).
Goldspink, C. (2000). Modelling social systems as complex: Towards a social simulation meta-model. Journal of Artificial Societies and Social Simulation vol. 3 no. 2 article 1. (http://www.soc.surrey.ac.uk/JASSS/3/2/1.html)
Gotts, N. M., J. G. Polhill, & A. N. R. Law. Agent-based simulation in the study of social dilemmas. Submitted to Artificial Intelligence Review.
Polhill, J. G., N. M. Gotts, & A. N. R. Law. (2001). Imitative versus nonimitative strategies in a land-use simulation. Cybernetics and Systems vol. 32 no. 1-2 pp. 285-307