Using trainable computing networks in the control of a physical system

We have applied certain computing networks to control a physical system. In our case the physical system is a reservoir network which is controlled by regulating the outflows of the reservoirs. At first we code the already existing control rules for the system into the structure of the computing network. The control rules define the regulation on the basis of the state of the system and the forecasted future inflows. After this the parameters of the computing network are fitted by an evolutionary and a direct-search optimization algorithms. Finally the optimized regulation rules are extracted from the network into a readable form. The optimized utility function consists of the effects of the regulation which are simulated using a stochastic model for the reservoir network system. The inputs of the model are a distribution of possible future inflows into the reservoirs and the the control rules for the regulation and the output is the expected value for the utility function. Compared to the traditional dynamic programming approach this simulation method allows virtually any utility function. In the utility function can even be considered such complex processes as ice accumulation, erosion and routing of flood wave. Another advantage of our approach is that the existing control rules for the system can be utilized by coding them into the computing network and after fitting the resulting optimized control rules can be extracted from the network and presented in a readable form.