Adaptive IoT systems in an open world setting (unknown unknowns) cannot be completely defined and realized during design-time, online learning to learn and improve the way that a system adapts during runtime is essential.

Current online learning solutions were devised for systems (such as cloud and web) that can tolerate slow convergence and thus require sufficient time to learn. This is no longer sufficient for the highly dynamic IoT systems setting, where each actuation and action may have an effect (even negative) in the environment. Thus, novel online learning mechanisms are required.

Online Learning enabler (enhanced reinforcement learning module taking into account the structure of the IoT systems' adaptation space)

As  an academic partner exploitation will include publications, offering training courses, as well as using the ENACT outcomes as part of graduate teaching.

Demonstrator, communications & publications, courses material

Prototype software (available as open source research demo)

Increased convergence of online learning

Expected TRL4 at the end of the project

The online learning enabler will empower IoT systems to self-adapt at runtime, even if the developers where not able to fully capture all potential future situations during design time.

Incorporating knowledge about the structure of the software systems's adaptation space makes the reinforement learning algorithms capable of exploiting knowledge that standard RL algortihms are not aware of.

IoT DevOps Engineer

Generic