I am currently reading Approximate Dynamic Programming by Warren Powell. The book comes out of work that Powell did with Schneider National, a company that operates a large fleet of trucks in the US.
I haven’t gone very far in the book, but it definitely seems very promising. The central premise that I can grasp is that it is fruitful to consider the quality of a model strictly in light of the decisions that the model is being used to make. This can be contrasted with typical surrogates like the root mean squared error or mutual information – those measures for model quality are largely derived using definitions of convergence towards total knowledge, and attempting to reach the state of total knowledge is often futile in reality.
In building practical models with practical approximations, choices in the early steps of an analysis often contain implicit knowledge about what the output from those steps is to be used for. To carry this process to its logical conclusion would be what I imagine approximate dynamic programming accomplishes – this way of dealing with ignorance trades off uncertainties in such way as to minimize the impact they have on decision quality, as measured by what you actually care about for the specific problem.