Continuous multi-task Bayesian Optimisation with correlation

This paper considers the problem of simultaneously identifying the optima for a (continuous or discrete) set of correlated tasks, where the performance of a particular input parameter on a particular task can only be estimated from (potentially noisy) samples. This has many applications, for example, identifying a stochastic algorithm's optimal parameter settings for various tasks described by continuous feature values. We adapt the framework of Bayesian Optimisation to this problem. We propose a general multi-task optimisation framework and two myopic sampling procedures that determine task and parameter values for sampling, in order to efficiently find the best parameter setting for all tasks simultaneously. We show experimentally that our methods are much more efficient than collecting information randomly, and also more efficient than two other Bayesian multi-task optimisation algorithms from the literature.