A constraints model was applied to a professional task that involved the regulation of a pressing-machine used to manufacture plastic products. The machine was devised so that visual and structural defects are avoided. This problem-solving model was formerly developed for classic puzzles such as the Towers of Hanoi or river-crossing problems. It is based on the idea that solving a problem is a compromise between three categories of constraints: goals based on knowledge about the situation, interpretations of objects and actions, and everyday learning heuristics. Constraints are rules that are defined on a set of possible actions, which allow some of them and prohibit the rest. They may be combined in such a way that restrictions are cumulated and a set of constraints is more restrictive. This may result in an impasse, a situation where each action is prohibited by at least one constraint. Resolving the impasse is obtained by the relaxation of the least preferred constraint, assuming that there is an ordering of constraints. Contrary to other problem-solving models, the constraints model makes it possible not only to simulate a protocol but also to make a diagnosis of the set of constraints which enables the protocol to be simulated at best. Thus, it provides an analysis of the problem-solving process at the individual level and makes clear the differences as well as the similarities in the way that problems are solved. First, the possible rules have to be extracted through a hand-made analysis of protocols and then expressed in the form of constraints. A set of algorithms is programmed which, through a series of steps, selects the best set of constraints using a measure of distance between the data obtained and the results of the simulation.
The task implies both that information is requested and that the parameters which regulate the pressing-machine are changed. Thus, two sets of decisions have to be considered and two decision spaces have to be defined, where: (i) a choice has to be made between an action on the device or an information request about which defects are present at that moment, or the curve of pressures which indicate the state of some of the parameters of the system; and (ii) a choice has to be made regarding which parameter to change and in what direction, when the decision has been taken to modify the system. The protocols of 13 regulators that solved a series of problems have been collected in an experimental setting, which simulates the interaction with the machine. For each of them it has been possible to find a set of rules, which accurately simulates behaviour. Beyond general heuristics, the analysis revealed three main strategies in dealing with the task. The simpler one does not take into account the information from the curve; it relies only on knowledge of the possible causes of defects and on trial-to-trial feedback. The second one relies only on the information provided by the curve in order to modify the parameters readable from the curve, and requests information about defects only when the curve looks normal. The third one always compares the information provided by the defects and the information provided by the curve. The results are discussed in relation with the types of strategies described in the literature for diagnostic tasks.