Hyperparallelepiped Based Loop Partitioning of Adaptive Dynamic Programming Structure

This paper focuses on an accelerating method for partitioning the loops in the structure of the adaptive dynamic programming (ADP). ADP contains critic-actor structure which involves the iterations of the value function. When the system needs to be stable, the value function generally needs to iterate thousands of times, the high computation burden which hinders the iterations will be generated. In order to reduce the computation burden, we introduce a hyperparallelepiped based loop partitioning (H-LP) method which splits the iterations of the value function and reduces the communication traffic calculated by the data footprint. The experiment results show that the computation performance will be enhanced when the H-LP method is introduced. The proposed method has an important practical significance.