3.2.2S
A REVIEW OF ANALYTICAL TECHNIQUES FOR DECISION SUPPORT

DJ PARSONS

Silsoe Research Institute, Wrest Park, Silsoe, Bedford, MK45 4HS, UK

Background and objective
The purpose of a decision support system is to acquire data, process or analyse it, and present relevant, timely information to a decision maker in a form that assists in making decisions. This may seem obvious, but it is probably true to say that most of the failures of decision support systems have resulted from neglecting some aspect of this definition. This paper will concentrate on techniques for converting data, particularly that produced by models, into specific assistance in decision making, while recognising that this is only part of the process. It will conclude with a case study of the approaches taken to supporting decisions on the control of fungal diseases in winter wheat as part of the DESSAC project.

Results and conclusions
There are many types of decision support systems, based on 4 main technologies: (i) management information systems which collect data and present it to growers to allow them to draw their own conclusions; (ii) descriptive or predictive models: for example disease simulation models that allow users to experiment with the effects of different treatments; (iii) normative or prescriptive models, that is, those that use models such as the ones mentioned above, but use the resulting information to try to derive the best course of action; (iv) expert systems, or other artificial intelligence approaches: these are also normative, but model the decision making process, rather than the underlying system.

The use of normative models is the subject of this paper. This approach - modelling a part of the system under consideration and using optimisation or a heuristic to derive a course of action - has been a feature of operational research since Wodd-War 11, and most of the techniques discussed come from that discipline. Two main approaches are considered: taking a predictive model of the system and applying an optimisation algorithm to it; and formulating the model specifically for optimisation.

The first approach is very effective for problems with relatively few continuous variables, but more difficuft in many practical problems that involve discrete or combinatorial variables, such as chemical selections. For these problems, some of the newer stochastic optimisation techniques, such as simulated annealing and genetic algorithms are more appropriate.

The second approach, uses mathematical programming techniques, of which the best-known is linear programming. This is very efficient for solving very large problems, and is especially useful at the level of whole-farm management. However, it requires the model to be constructed in a very specific form. Newer MP methods allow the use of integer vadables, limited nonlinearity, multiple criteria and stochastic elements, but are still not suitable for all types of models. One technique of this type, dynamic programming, is particularly useful for sequential decisions on stochastic problems. Several agricultural problems have been formulated in this way, including autumn cultivation decisions, litigation and weed control. It is a powerful method for solving such problems and results in a complete policy giving the optimum action for every state of the system. However, although it is much more efficient than exhaustive search methods, it rapidly becomes impractical when the number of state variables is increased.

The wheat disease control module within the DESSAC project provides an interesting case study. A simulation model of crop development, fungal disease and the effect of spraying is being developed, and a technique for finding the best spray programmes is required. A number of features make this difficult: (i) the state of the system cannot be described by a small, fixed set of state variables (making dynamic programming inappropriate); (ii) it includes integer and combinatorial variables (such as chemical selection); (iii) the users require a list of several 'good' programmes, not a single optimum These characteristics led to the choice of a genetic algorithm as the method of solution. This has now been applied to tactical (spring) decision making with some success.