Contextual bandits algorithms for reconfigurable hardware accelerators

Reconfigurable processing cores for IoT and edge computing applications are emerging topics to calibrate costs, energy consumption and area occupation with performance and reliability on Commercial Off the Shelf (COTS) devices. This work analyzes how to take advantage of Machine Learning to potentially automate the reconfiguration process of a hardware accelerator inside the Klessydra Vector Coprocessor Unit (VCU), choosing the best configuration according to the workload. The problem is modeled with a contextual bandits approach using the Linear UCB algorithms and validated with offline Python simulations.