A digital twin is typically considered a combination of an inferenc, or data-assimilation, problem to understand what is happening now, and an optimization problem to define the best action for the future. However, by formulating the optimization problem also as an inference problem, the action becomes a probabilistic decision problem, pushing the action, and hence the digital twin to the next level. In this talk, the new formulation is explained and demonstrated on (most likely still) a problem of limited complexity.