On the Use of Reinforcement Learning for Real-Time System Design and Refactoring

Abstract

One crucial issue during the design of real-time embedded systems is the deployment of tasks on distributed processors. Indeed, due to the constrained resources of these systems and their real-time aspect, the allocation model must be valid and optimal. This design phase is generally time-consuming and difficult, especially when the design decisions need to be frequently updated (i.e., refactoring). To address this problem, we provide in this paper an optimization model based on the reinforcement learning (RL) approach. The proposed model produces the optimal deployment model, ensuring that timing properties are respected while minimizing the number of active processors. After a refactoring request, the generation of the new solution in the proposed RL model is greatly based on the initial one. Indeed, just a minor part of the solution has to be updated, which is beneficial since it results in a shorter generation time. The efficiency of the proposed model is demonstrated through a case study and performance evaluation.