Aggregation in Dynamic Programming

Reducing the size of a dynamic program through state aggregation can significantly reduce both the data and the computation time required to solve a problem. We develop a new algorithm that combines state aggregation and disaggregation steps within a single-pass procedure. The solution obtained is automatically feasible for the original problem. By exploiting general conditions on the aggregate structure, we develop bounds for assessing the error from optimality introduced by the aggregation, and we illustrate with an application in infinite horizon optimization.