Support for using portfolios of acquisition functions, not just a single acquisition function.
There is some support that this yields a superior performance than any single function (if computational budget allows it...); for example, see paper "An Entropy Search Portfolio for Bayesian Optimization" (https://arxiv.org/abs/1406.4625)
Maybe a nice feature to also allow us to frame Thompson sampling as a "classical" BO algorithm, where the acquisition function is a GP sample. The porfolio of acquisitions would be different samples?
Support for using portfolios of acquisition functions, not just a single acquisition function.
There is some support that this yields a superior performance than any single function (if computational budget allows it...); for example, see paper "An Entropy Search Portfolio for Bayesian Optimization" (https://arxiv.org/abs/1406.4625)