Abstract
Spatially and temporally explicit simulation modelling of natural resource management systems provides a framework to draw together in a mathematically unambiguous manner a wealth of information, and, as such, allows rigorous testing of hypotheses of how such systems can be changed, without the time, expense and moral implications of altering a real system. In recent years, a large number of integrated assessment models linking the human and biophysical components of particular systems have been developed to address this need, but in many of these models the human dimension is based on economic cost-benefit principles that attempt to optimise use of resources such as capital or labour to maximise a particular output. Limitations to these approaches are that they are structured to represent an equilibrium when production has stabilised, they presuppose a 'goal' of the system, and do not adequately consider the microdecisions being made by the various actors within it. Agent-based modelling (ABM) is an approach that has been receiving attention in recent years as a way of linking the biophysical and socioeconomic characteristics of a system, and which provides a way of addressing these limitations. ABM has aroused the interest of environmental modellers, mainly because it offers a way of incorporating the influence of human decision-making on the environment in a mechanistic and spatially explicit way, taking into account social interaction, adaptation, and multiple scales of decision-making. Several such models are now beginning to appear, many of which involve the grafting of an ABM representing a number of households onto a cellular automata 'landscape', with each agent being linked in some way to the cells over which it has influence. Apart from changes in actual land cover, however, these models generally treat the landscape as a relatively static entity, and do not simulate processes such as soil water and nutrient dynamics. The ones that do include such processes, do so somewhat simplistically. There is a need, therefore, to integrate dynamic biophysical simulation models with these emerging agent-based social simulation models. Different approaches to integrating such models are recognised - one such scheme refers to 'loosely-coupled', 'closely-coupled', and 'fully integrated' levels of integration. Loose- and closely-coupled models exchange driving variables between them, with closely-coupled models sharing common subprocesses, meaning that temporal and spatial scales may be determined by the original (sub-)models being coupled together. By contrast, in fully integrated models, these scales are dictated by the processes being simulated. It is our view that it is necessary to focus on the fully integrated level in developing models to adequately understand the behaviour of managed ecosystems. We discuss an agent architecture that allows agents to communicate regardless of the programming language used - each agent should have a translation module that translates incoming messages and triggers the appropriate internal response, and a conversation module which checks ingoing and outgoing messages, and manages communications between multiple agents. The overall system should be coordinated by manager and router agents to ensure the provision of global information and correct delivery of communications between agents, respectively. To link this to different sub-models of biophysical processes, a limited number of common properties of the sub-models are required: (a) each sub-model must have the ability to advance one time-step on request, (b) it should be able to save the states of all its variables at the end of each time-step on request, and be able to reload these later, also on request, (c) it must be able to respond to predefined message requests for information, and (d) the calculation of rates of change of its state variables must be separate from the updating of those state variables, with both operations being carried out on request. There is a danger that such models become too complex - it is suggested that the best way forward may be to take a simple framework as the starting point, and incorporate additional detail as necessary to describe the processes of interest.