Bekæmpelsesmiddelforskning fra Miljøstyrelsen, 105 – Ukrudtsstriglingens effekter på dyr, planter og ressourceforbrug – Bilag C - Validation procedures for ALMaSS, and other complex adaptive systems

Ukrudtsstriglingens effekter på dyr, planter og ressourceforbrug

Bilag C

1 Validation procedures for ALMaSS, and other complex adaptive systems

ALMaSS is an agent-based model (ABM), or complex adaptive system (CAS). It is an entirely different class of model to traditional population models which are usually based upon simple equations and relationships. Agent-based complex systems such as ALMaSS are dynamic networks of many interacting entities mimicking the processes and interactions we see in the real world. Whilst unlike statistical modelling there is no general framework for designing, testing, and analyzing bottom-up models as yet established, recent advances in ecological modelling have come together in a general strategy called pattern-oriented modelling.

What is pattern-oriented modelling? The patterns refer to system properties that we see around us, e.g. population growth curves, spatial distribution of organisms, individual developmental rates, or any other measurable characteristic of a system resulting from the interaction of agents. Patterns are therefore the defining characteristics of the modelled system.

The test of whether a CAS model is sufficient to be able to explain the cause of dynamics of the system is whether it can generate enough sufficiently accurate patterns. If it can do this then the mechanisms that are used to build up the model are considered to be sufficient to generate system behaviour akin to that seen in the real world system. In other words if enough of the basic mechanisms are incorporated so that the model system responds to changes like the real world system, then sufficient confidence can be generated in the model to consider its use as a predictive tool.

There are two kinds of pattern that must be evaluated, basic patterns and complex patterns. The complex patterns are what are termed ‘emergent properties’ and are the product of interactions between agents in the model. The basic patterns are features of the behaviour of the model that are directly programmed in. Hence checking these requires a range of trivial although laborious tests.

Checking the basic patterns are more or less checks on the system to see that all is functioning as intended. In ALMaSS this includes checks on weather input, landscape structure, vegetation growth, crop allocation, and not least crop or other habitat management. All these basic model parts need to be checked to see that they are producing the correct patterns. The same process is also carried out for the animal models. All kinds of individual behaviours need to be checked to see if they function correctly. These kinds of behaviours are those that are part of the processes directly programmed into the model and are not emergent properties. These basic checks can be thought of as a complex debugging process, i.e. afterwards the individual sub-components of the model should behave according to expectations (e.g. a farmer should manage his crop following the plan provided).

Once all the basic patterns are correct the interactions between these results in emergent properties, i.e. complex patterns that are not directly programmed responses. Relatively few complex patterns are required for successful analysis since they are a function of a great many model components or mechanisms. Probably the simplest and first of these is the plausibility criteria. In ALMaSS where we are simulating animal species this usually consists of an ecologist’s evaluation of the behaviour of the model animals, i.e. do they behave like the real thing? Subsequently numerical comparisons with observed patterns are utilized. These patterns, such as changes in animal population numbers with time, are a result of integrating all factors affecting growth, mortality, reproduction and dispersal across spatially and temporally heterogeneous landscapes affected by agricultural management.

If the model can predict a number of emergent patterns simultaneously from the same set of inputs (e.g. spatial distribution of animals at the same time as their developmental rates), then confidence in the model grows. As we increase the numbers of patterns that are used to compare model outputs to, we reduce the set of potential input parameter values (parameter space) that can be used. Why is this? The reason is that each new pattern that is added requires a certain specific set of parameter values to achieve a good match, as new patterns are used these new values much be selected from the set of values used to match the preceding patterns. The result is therefore an ever-dwindling range of possible parameter values that can achieve a match to all patterns; hence adding further patterns adds further limitations to the inputs. If sufficient real world data were available to add sufficient new patterns, then the continual reduction in input parameter space will eventually lead to a situation where deviation from a distinct set of parameters leads to failure to simulate all patterns. At this point the model cannot be improved and not only is the model structure (interactions between agents) well tested, but the inputs are also narrowed down, reducing uncertainty and providing a powerful predictive capability.