Resource Allocation via Bayesian Optimization: an Efficient Alternative to Semi-Bandit Feedback

Although optimal resource allocation is a well-known and studied problem, the recent technological innovations are bringing to light new specificities and issues. Some relevant real-life applications are the optimal management of cloud/high-performance computing resources, and the optimal budget allocation for multi-channel marketing. Recent formulations have led to the definition of the Semi-Bandit Feedback approach, that is the reference method in these emerging real-life settings. In this paper we propose a novel approach, extending the Bayesian Optimization framework to specifically deal with the resource allocation problem, and finally resulting more efficient than Semi-Bandit Feedback. Moreover, the proposed approach can also deal with specific (real-life) settings that cannot be covered by Semi-Bandit Feedback. We have validated our approach on (i) the case study reported in the original paper proposing Semi-Bandit Feedback, (ii) a multi-channel marketing application, and (iii) the optimal mix of water sources in water distribution networks.