A selective improvement technique for fastening Neuro-Dynamic Programming in Water Resources Network Management

Implementing integrated water resources management paradigm is a hard task with respect to modelling and computational aspects it involves, but its centrality in effective water management suggests to address these aspects by strengthening the efficiency of the solution techniques, instead of weakening the integration requirements. When dealing with networks of water resources SDP provides an effective solution methodology as it amplifies its efficiency boost. On the other hand, SDP suffers from the so-called “curse of dimensionality”, that rapidly leads to the problem intractability. Neuro-Dynamic Programming can sensibly mitigate both of these drawbacks by approximating the Bellman functions with ANN. NDP shows to be considerably slowed by the ANN training phase. To overcome this limit a new training architecture has been developed, taking advantage of the NDP peculiarity.