Using human-machine dialogue to refine generalisation evaluation function

Patrick Taillandier (1) , Julien Gaffuri (2)

1 UMMISCO - Unité de modélisation mathématique et informatique des systèmes complexes [Bondy]
2 COGIT - Conception Objet et Généralisation de l'Information Topographique

More and more geographical data producers use automated generalisation to produce their data. Indeed, the spreading of artificial intelligence techniques has allowed an important improvement of the generalisation process automation. A classic approach consists in formalising generalisation as an optimisation problem: the goal is to find a state of the data that maximises a function. All these methods use a function that is supposed to assess the generalisation state of the data, according to the user need. We propose to call this function "evaluation function" (the expressions "utility function", "fitness function", "energy" or even "satisfaction" are also often used). A key issue of this approach concerns the design of this evaluation function. Indeed, in order to get good results, such systems have to know what it is searching, i.e. what a good generalisation of the input data is. Unfortunately, designing such a function remains a difficult task. Indeed, while the final user of the generalised data can easily describe his need in natural language, it is often far more difficult for him to express his expectations in a formal language that can be used by generalisation systems. This problem is particularly complex when numerous measures are used to characterise the quality of a generalisation and when no simple links between these measures values and the solution quality can be found. In this paper, we propose an approach dedicated to the design of generalisation evaluation functions. An evaluation function previously designed by a user is improved through a dialogue between this user and the generalisation system. The idea is to collect user preferences by letting the user compare different generalisation results for a same object (or group of objects).

Domaines: Apprentissage [cs.LG]; Interface homme-machine [cs.HC]

Patrick Taillandier, Julien Gaffuri. Using human-machine dialogue to refine generalisation evaluation function. Conference of the International Cartographic Association, 2011, Paris, France. ⟨hal-00688425⟩