FORMATION OF THE STRUCTURE OF A NEURON NETWORK THROUGH DECOMPOSITION OF THE INITIAL TASK ON A PARTICULAR EXAMPLE OF THE ROBOT-MANIPULATOR MANAGING PROBLEM

Currently, neural networks are widely used in the construction of many systems in which the development of an algorithm is a serious problem. However, neural networks are considered as a «black box», and its structure is in practice obtained by carrying out a lot of experiments and choosing the best option available. The article proposes a method of forming a neural network structure on the basis of the original problem decomposition, resulting in a set of states the system can be in, and the signs of the state changes. Next, it is proposed to construct a finite automaton whose structure is the main one for the realization of a neural network. Building of a neural network consists of three stages: 1) the implementation of the part for determining the signs of transitions between states; 2) the implementation of the part for determining the state of the system; 3) the implementation of a neural network for each state for generating the output signals and combining the output signals from different states. The resulting neural network allows for greater observability and the possibility of «debugging», as well as greatly simplifies the learning process by using simpler types of neural networks and problems being solved

Authors: А. А. Voevoda, D. O. Romannikov

Direction: Informatics and Computer Technologies

Keywords: Neural networks, neural network structure, reinforcement training, automation, finite state machine, robot manipulator control

---

View full article