Evolution Strategies with an RBM-Based Meta-Model

摘要

Evolution strategies have been demonstrated to offer a state-of-the-art performance on different optimisation problems. The efficiency of the algorithm largely depends on its ability to build an adequate meta-model of the function being optimised. This paper proposes a novel algorithm RBM-ES that utilises a computationally efficient restricted Boltzmann machine for maintaining the meta-model. We demonstrate that our algorithm is able to adapt its model to complex multidimensional landscapes. Furthermore, we compare the proposed algorithm to state-of the art algorithms such as CMA-ES on different tasks and demonstrate that the RBM-ES can achieve good performance.