Machine Learning (ML) is concerned with the development of techniques allowing the computer to acquire knowledge. In our work, a ML component resolves all questions related to the system’s selection of media and modalities with reference to the user’s situation. This ML component uses a priori training sets which contain records of scenarios. Every scenario record is composed of a pre-condition scenario (i.e. the user context), and its corresponding post-condition scenario (i.e. the media and modalities that are appropriate for such context).  

Given a certain context, the media and modalities listed in the post-condition scenario are set for activation. A problem arises when a media or modality is found missing or defective which could potentially cause the system to stall or to crash. Our system uses its acquired knowledge to find a replacement to the defective component. To do so, the ML agent consults its knowledge database which contains a list of replacements to a failed component. If the list is empty or when the selected device/modality and all its replacements have all failed, the ML system is trained; each training yields one device included in the list. The more the ML system is trained, the more resilient the system becomes over components failure.

This paper demonstrates the design of a fault-tolerant ubiquitous MM computing system. The requirements analysis is undertaken by considering the quality attributes desired by different stakeholders. We use the attribute-driven design method in the requirement analysis and Architecture Tradeoffs Analysis Method in evaluating the system architecture.