Evaluation of the Learning Classifier System XCS for SoC run-time control

Abstract

In this paper, we evaluate the feasibility of using the learning classifier XCS to control a System-on-Chip. Increasing number of transistors and process variation make it difficult for a chip designer to foresee all possible run-time conditions. Post-poning some decisions from design time to run time alleviates the designer’s life and allows shorter time-to-market. In this paper, we evaluate if XCS can take these run-time decisions on a processor with four cores. The evaluation shows that XCS can find optimal operating points, even in changed environments or with changed reward functions. This even works, though limited, without the genetic algorithm the XCS uses internally. The results motivate us to continue the evaluation for more complex setups.