Artelys Knitro is now available on ARM and opens up to the embedded world

For instance, a vision system that detects defects on a production line may have learnt from a set of data comprised of both compliant and faulty products. In practice this training involves the adjustment of many parameters in order to minimize the error rate. This often means solving a nonlinear optimization problem.

 

During the training phase of the model, many parameters are chosen in order to minimize a nonlinear objective function comprised of both the error and a regularization term.

 

In many cases, the training phase is carried out in the cloud, and the trained model is used to perform inference on the device. However, this process is slow because of the time needed to transmit the data and model updates between the device and the server. In situations where the device is expected to learn and respond to the new scenario instantly, there will be a need for local model updates. Local training also addresses the important issue of privacy, as the data does not need to leave the device.

Self-driving cars integrate machine learning and optimization in a complimentary way. The first enable the system to learn how to identify static and moving objects, as well as signalization.

 

Then the vehicle must compute a trajectory towards the desired location that maximizes safety and comfort while minimizing energy consumption and travel time, taking into account obstacles, signalization, road shape… This is a nonlinear optimization problem. And there are many more examples of embedded optimization such as drone stabilization or rocket landing!