Neural approximators for the solution of decentralized optimal control problems

Abstract

There are many situations, in engineering and economic systems, where several decision makers (DMs), sharing different information patterns, cooperate to the accomplishment of a common goal. We address an approximate technique consisting in constraining the control functions to have a fixed structure (we chose feedforward neural networks). We are then able to obtain solutions that approximate the optimal ones within any desired degree of accuracy under very general conditions. Such a technique has proved to be effective in non-LQG classical optimal control and in team problems not solvable analytically.