MUMBO: MUlti-task Max-Value Bayesian Optimization

We propose MUMBO, the first high-performing yet computationally efficient acquisition function for multi-task Bayesian optimization. Here, challenge is to perform optimization by evaluating low-cost functions somehow related our true target function. This a broad class of problems including popular task multi-fidelity However, while information-theoretic are known provide state-of-the-art optimization, existing implementations scenarios have prohibitive computational requirements. Previous therefore been suitable only with both low-dimensional parameter spaces and query costs sufficiently large overshadow very significant overheads. In this work, we derive novel version entropy search, delivering robust performance low overheads across classic challenges hyper-parameter tuning. MUMBO scalable efficient, allowing be deployed in rich fidelity spaces.