Small surface-mounted catalytic cluster tend to be fluxional and have many low-energy isomers. For realistic modeling that would approach conditions of catalysis, it is therefore required to model cluster catalysts as a statistical mechanical ensemble, with properties being calculated as ensemble averages. We will present our methods developed for this purpose: the GPU-accelerated Neural Network global optimizations, and statistical modeling of cluster sintering, reactivity, and selectivity. The focus application will be Pt clusters on oxides used for catalytic dehydrogenation of alkanes. Pure Pt clusters deactivate rapidly via sintering and coke deposition. We find that alloying Pt clusters with B stabilizes the clusters against both mechanisms of deactivation. The predicted anti-coking effect of boration is then confirmed experimentally. We emphasize that this prediction can be borne only within the introduced ensemble description of cluster catalysts, whereas the traditional approaches to modeling cluster catalysts yield completely opposite predictions.