Full Implementation of an Estimation of Distribution Algorithm on a GPU

We submit an implementation of an Estimation of Distribution Algorithm – specifically a variant of the Bayesian Optimisation Algorithm (BOA) – using GPGPU. Every aspect of the algorithm is executed on the device, and it makes effective of use multiple GPU devices in a single machine. As for other EDAs, our implementation is generic in that it may be applied to any problem for which solutions may be represented as binary strings. For the purpose of this paper, we apply it to a particular problem known to be difficult for metaheuristic algorithms due to high interdependency between variables: finding the lowest energy state of an Ising Spin Glass. We show that our GPU implementation demonstrates a speedup in excess of 80x compared with an equivalent CPU implementation. To our knowledge, this is the first EDA to be implemented fully on the GPU.