arXiv:2406.03616v1 Announce Sort: new
Summary: Novelty search (NS) refers to a category of exploration algorithms that robotically uncover numerous system behaviors by way of simulations or experiments. Systematically acquiring numerous outcomes is a key part in lots of real-world design issues reminiscent of materials and drug discovery, neural structure search, reinforcement studying, and robotic navigation. Because the relationship between the inputs and outputs (i.e., behaviors) of those complicated techniques is often not obtainable in closed type, NS requires a black-box perspective. Consequently, common NS algorithms depend on evolutionary optimization and different meta-heuristics that require intensive sampling of the enter area, which is impractical when the system is pricey to guage. We suggest a Bayesian optimization impressed algorithm for sample-efficient NS that’s particularly designed for such costly black-box techniques. Our strategy fashions the input-to-behavior mapping with multi-output Gaussian processes (MOGP) and selects the following level to guage by maximizing a novelty metric that is determined by a posterior pattern drawn from the MOGP that promotes each exploration and exploitation. By leveraging advances in environment friendly posterior sampling and high-dimensional Gaussian course of modeling, we talk about how our strategy may be made scalable with respect to each quantity of information and variety of inputs. We check our strategy on ten artificial benchmark issues and eight real-world issues (with as much as 2133 inputs) together with new purposes reminiscent of discovery of numerous steel natural frameworks to be used in clear vitality know-how. We present that our strategy tremendously outperforms current NS algorithms by discovering considerably bigger units of numerous behaviors underneath restricted pattern budgets.
Supply hyperlink