Using neural networks for solving game tasks on the example of the problem of searching the way in the labyrinth

The article deals with the solution of game problems by the example of finding a path in a labyrinth with the help of a neural network. Such a task can be solved by one of the existing methods of training with confirmation, but the solutions have several drawbacks, which include, for example, a long learning time for the “classical” training method with confirmation. Analysis of these algorithms is presented, including a training method with confirmation based on Monte Carlo search method in the game tree. There has been described a variant of setting the initial problem, in which the input consisting of only one value — the sign indicating that the previous step was successful or deadlock — is delivered to the neural network, instead of the entire field of the labyrinth. A variant of solving the problem is posed by the synthesis of a neural network based on a preliminary synthesized algorithm. A neural network obtained is constructed from an informal description of the algorithm; it allows searching for a path in the labyrinth. The main idea used in the formation of a neural network is that the detected or explored path is marked in the neural network by the weights of the units, the dead-end path — by the negative weights, and the unexplored path — by the zero weights. When a robot occurs in a dead end, the process of finding the path is restarted from the initial state, and the dead end state is marked in the labyrinth as unattainable. Several stages of the path search process in the labyrinth illustrating the operation of the synthesized neural network are presented. There has been given a comparative analysis of the solution obtained and of the existing methods of training with confirmation.