Abstract
In many engineering problems including control problems, optimization of the policy for multiple conflicting criteria is required. However this is very challenging if there exist noise, which may be input dependent, and/or the restriction in the number of evaluations, which is induced in the case where the experiments are expensive in time and/or money. This paper presents a multiobjective optimization (MOO) algorithm for expensive-to-evaluate noisy functions. By incorporating a heteroscedastic Gaussian process regression method as well as standard Gaussian process regression, the algorithm creates suitable surrogate functions from noisy samples and finds the point to be observed at the next step. This algorithm is compared against an existing MOO algorithm, and then applied to optimize the sidewinding gait of a snake robot.
