reflecting on past work – offline model-based optimisation

reflecting on past work -- offline model-based optimisation $$\newcommand{\argmax}{\text{argmax}}$$ $$\newcommand{\ft}{f_{\theta}}$$ $$\newcommand{\ftrain}{f_{\text{train}}}$$ $$\newcommand{\fvalid}{f_{\text{valid}}}$$ $$\newcommand{\ftest}{f_{\text{test}}}$$ $$\newcommand{\fphi}{f_{\phi}}$$ $$\newcommand{\ftt}{f_{\theta}}$$ $$\newcommand{\ds}{\mathcal{D}}$$ $$\newcommand{\pt}{p_{\theta}}$$ $$\newcommand{\ptnew}{\widehat{p_{\theta}}}$$ $$\newcommand{\ptrain}{p_\text{train}}$$ $$\newcommand{\pvalid}{p_\text{valid}}$$ $$\newcommand{\dtrain}{\mathcal{D}_{\text{train}}}$$ $$\newcommand{\dvalid}{\mathcal{D}_{\text{valid}}}$$ $$\newcommand{\dtest}{\mathcal{D}_{\text{test}}}$$ $$\newcommand{\drest}{\mathcal{D}_{\text{rest}}}$$ Table of Contents 1. intro - what is model-based optimisation? 1.1. formalising offline mbo 1.2. reward-based extrapolation 1.3. ‼️ why evaluation is difficult (and misunderstood) 1.4. training, validation, and testing 2. last year's work 2.1. ranking validation…