The work presented in this paper is in the field of Technology for Human Learning (TEL), specifically that of engineering in the context of TEL-type approach "Learning Design". This approach proposes to build ILE from the formal description of a learning activity. It assumes the existence of a common modeling language called EML (Educational Modelling Language) and an engine capable of interpreting this language. LDL is the language on which we worked in conjunction with the LDI infrastructure including a motor interpretation of LDL. The EML is used to produce a scenario, a formal model of learning activity. The ILE serving to support the conduct of the activity described in the scenario is then constructed semi-automatically in the infrastructure associated with the language using the following process: the script is created during a design phase; it is instantiated and deployed on a platform of services in an operational phase (selection of participants in the activity, role assignment, choice of resources and services); instantiated and deployed scenario is supported by the engine will interpret it to ensure its implementation. In this context, the activity takes place in accordance with what was specified in the script. It is impossible to predict in advance all that can occur in an activity, the activities are by nature unpredictable. Unforeseen situations can occur and lead to disturbances in the activity, or even blocks. It then becomes important to provide the means to break the deadlock. The teacher may also want to exploit a situation or opportunity by altering the activity in progress. This is the problem that is addressed in this thesis: to provide the means to adapt an activity "on the fly", that is to say while running, so you can not handle a situation provided and continue the activity. The proposal we make is based on the differentiation between data convened in each of the three phases of construction of the ILE: design, operationalization and implementation. We exhibit a model for each phase, which organizes the data and positions to each other. Adapt an activity "on the fly" then returns to modify these models in different situations to deal with. Some require the modification of one of its models, while others lead to propagate changes from one model to another. We consider adaptation "on the fly" as a separate activity conducted in parallel with the learning activity by a human supervisor has an adequate environment to observe the activity, identify potential problems and be remedied by intervention in the learning activity by modifying the templates that specify. To develop tools to support the change and integrate them into the LDI infrastructure, we have used the techniques of Model Driven Engineering. Handled in these models are tools and data in its own right tools and made no offer more flexibility and abstraction. The models are then used as levers to achieve and change the data targeted by the adaptation.
