Energy Aware Strategy for Discrete Event Systems using Inhibitor P-Time Petri nets and Deep Reinforcement Learning
Résumé
Energy considerations become a critical issue for
modern man-made systems and the need for efficient ecoresponsible
solutions ensuring energy savings is crucial. This
paper develops an energy aware method allowing to optimize
the energy consumption of data centers systems thanks to
the introduction of Inhibitor P-Time Petri nets (IP-TPN) and
Deep Reinforcement Learning techniques. Indeed, thanks to a
schedulability analysis method and being given an energy cost
function, the global energy consumed for a particular behavior
of the system considered can be computed. Furthermore, Petri
nets are a good framework for deep reinforcement learning, on
the one hand because the cost function we introduce in this
paper will naturally train an agent to minimize the total cost
according to random inputs, and on the other hand because
the observation space containing the tokens as well as their
age in the square is high dimensional, which makes traditional
algorithms less efficient than deep reinforcement learning.