Managing product inherent variance during treatment

The natural variance of agricultural product parameters complicates recipe planning for product treatment, i.e. the process of transforming a product batch from its initial state to a prespecified final state. For a specific product P, recipes are currently composed by human experts on the basis of heuristic matches between product state and recipe features. This approach makes use of standard recipes, that do not sufficiently reflect inherent differences between batches. Improvement of the recipe design process requires three problems to be solved: (1) assessment of the initial product state, (2) fixation of the recipe requirements and (3) design of a treatment recipe. To objectively assess the initial product state, additional measurement of a specific parameter is required. This parameter varies substantially between batches, requiring large measurement samples. Without objective assessment, however, automated determination of the recipe requirements and recipe design is not possible. This paper describes a procedure to get an objective initial state assessment, and presents a Product Treatment Support System that takes an initial state assessment, and performs the process of recipe design. Artificial intelligence (AI) techniques are applied at three points in the process. Induction of decision trees is used to determine rules that are understandable to experts and that select products most suitable for state assessment. Neural networks are applied to transform the assessment of the initial state into the overall requirements of the recipe. Finally, the actual recipe is derived by means of constraint satisfaction