high performance computing on graphics processing units: hgpu.org

We develop the first parallel algorithm for Coalition Structure Generation (CSG), which is central to many multi-agent systems applications. Our approach involves distributing the key steps of a dynamic programming approach to CSG across computational nodes on a Graphics Processing Unit (GPU) such that each of the thousands of threads of computation can be used to perform small computations that speed up the overall process. In so doing, we solve important challenges that arise in solving combinatorial optimisation problems on GPUs such as the efficient allocation of memory and computational threads to every step of the algorithm. In our empirical evaluations on a standard GPU, our results show an improvement of orders of magnitude over current dynamic programming approaches with an ever increasing divergence between the CPU and GPU-based algorithms in terms of growth. Thus, our algorithm is able to solve the CSG problem for 29 agents in one hour and thirty minutes as opposed to three days for the current state of the art dynamic programming algorithms.