METHOD OF CONSTRUCTION OF A NEUROREGULATOR MODEL WHEN OPTIMIZING THE CONTROL STRUCTURE OF A TECHNOLOGICAL CYCLE

In this paper authors present the results of a research that had a purpose to develop a method of constructing a neuroregulator model for the case of optimization of the control structure of a technological cycle. The method's implementation is based upon the automation of a production process when a physical controller, that operates the technological process according to a given program, is present. In order to achieve this goal, the artificial neural network approaches were implemented to create a mathematical model of the neuroregulator. The mathematical model of the neuroregulator is based on a physical prototype, and the procedure of a real-time control synthesis (adaptive control) is based on recurrent neural network training. The neural network architecture includes LSTM blocks, which are capable of storing information for long periods of time. A method is proposed for constructing a neuroregulator model for control of a production cycle when solving the task of the optimal trajectory finding on the phase plane of the technological cycle states. In the considered task of the optimal trajectory finding the mathematical model of the neuroregulator receives at each moment of time information about the current system state, the adjacent system states and the movement direction on the phase plane of states. Movement direction is determined by the given control optimization criteria. Based on the research results it was found that recurrent networks with LSTM modules can be used successfully as an approximator for the agent's Q-function to solve the given problem when the partially observed region of system states has a complex structure. The choice of the method of adaptation to the control actions and the external environmental disturbances proposed in the paper satisfies the requirements for the adatation process performance, as well as the requierments for the control processes quality, when there is lack of information about the nature of random control disturbances. The experimental environment, as well as the neural network models was implemented using the Python programming language with TensorFlow library.

1. Maksimej I.V., Smorodin V.S., Demidenko O.M. [Development of simulation models of complex technical systems]. Gomel: GGU im. F. Skoriny; 2014. (in Russ.)

2. Hochreiter S., Schmidhuber J. Long short-term memory. Neural Computation. 1997;9(8):1735-1780. doi:10.1162/neco.1997.9.8.1735.

3. Sutton R.S., Barto A.G. Reinforcement Learning: An Introduction. Cambridge: The MIT Press; 1998.

4. Mnih V., Kavukcuoglu K., Silver D., Rusu A., Veness J., Bellemare M., Graves A., Riedmiller M., Fidjeland A., Ostrovski G., Petersen S., Beattie C., Sadik A., Antonoglou I., King H., Kumaran D., Wiestra D., Legg S., Hassabis D. Human-level control through deep reinforcement learning. Nature 2015;518(7540):529-533. DOI:10.1038/nature14236.

5. Toffoli T., Margolus N. Cellular Automata Machines: A New Environment for Modeling. Cambridge: The MIT Press; 1987.