Parallel Tracking and Reconstruction of States in Heuristic Optimization Systems on GPUs

Abstract

Modern heuristic optimization systems leverage the parallel processing power of Graphics Processing Units (GPUs). Many states are maintained and evaluated in parallel to improve runtime by orders of magnitudes in comparison to purely CPUbased approaches. A well known example is the parallel Monte Carlo tree search, which is often used in combination with more advanced machine-learning methods these days. However, all approaches require different optimization states in memory to update or manipulate variables and observe their behavior over time. Large real-world problems often require a large number of states that are typically limited by the amount of available memory. This is particularly challenging in cases in which older states (that are not currently being evaluated) are still required for backtracking purposes. In this paper, we propose a new general high-level approach to track and reconstruct states in the scope of heuristic optimization systems on GPUs. Our method has a considerably lower memory consumption compared to traditional approaches and scales well with the complexity of the optimization problem.