The use of natural or sketch based interfaces is not extended at all due to the problems they present at the moment, like the lack of robustness, repeatability and reliability, which are not guaranteed in this kind of user interfaces. The actual recognizers used on line in interactive applications neither offer ideal solutions. They use to be rigid and the success ratio in the classification decreases when the number of gestures or symbols admitted by the system increases. This kind of interfaces are much convenient, so they could support the initial stages of design (conceptual design), in which the traditional pen and paper continues being used. On the contrary, the modern tools don’t improve the functionality of these traditional tools. So the idea is to create an interface that permits to draw parametric sketches on a computer and later use the sketches in CAD applications to create geometry, which is the principal aim of the thesis.
In the present thesis a reliable and robust sketch recognition paradigm that supports the user in the first conceptual design stages has been designed. In last term, this has allowed to develop a natural (it permits the spontaneous generation of ideas on paper), adaptative (it permits to define imprecise and incomplete geometry) and transparent (the user can change his activity without previous advertisement –no menus available- and his intention is captured) caligraphic interface.
In order to implement the designed interface, a multi-agent platform has been selected. It has been verified that agent based systems are valid for applications that need knowledge guided rules, where the context information is important to take decisions, and these decisions can be changed in terms of different possibilities, offering the user the possibility to draw freely no matter what he sketches, the number of strokes or the order they are introduced.
Besides, the interface has been endowed with a modular structure, permitting to add new gestures / symbols to the dictionary with the minimal intrusion in the designed system and making the system more flexible.
There are several innovative contributions that have been implemented on the interface, like the inclusion to the recognizer of the “interspersing” (the act of interrupting an action and continuing later in the same point), the “overtracing” (superimposed strokes for creating geometry imitating the artistic sketching on paper), and the automatic change mode (it’s not necessary to use menus or buttons for changing between the “geometry”, “command” or “introduction of restrictions” or another sketching symbols) recognizing the user’s intention automatically.
Besides, concerning the recognizer, a new segmentation method of the sketch has been developed and implemented. In the proposed method, apart from the vertex corners, the tangents corners are detected with success ratios close to 100%, what supposes an improvement over the segmentation methods that can be found nowadays in literature.
Finally, a frame for the optimization that permits the automated training of the system has been used. This training is being carried out by processes off line in the majority of applications that need of this previous workout (consuming long time). The developed optimization in this thesis permits the automated initial adjustment of the system parameters in order to reach the best possible solution. 
Summarizing, the developments accomplished in this thesis include the intensive knowledge of the image analysis techniques, segmentation techniques, agent-based systems, classifiers / recognizers, and optimization techniques.