Implementation of Q — Learning algorithm for solving maze problem

Machine learning is very important in several fields ranging from control systems to data mining. This paper presents Q — Learning implementation for abstract graph models with maze solving (finding the trajectory out of the maze) taken as example of graph problem. The paper consists of conversion of maze matrix to Q — Learning reward matrix, and also the implementation of Q — Learning algorithm for the reward matrix (similar to minimizing criteria matrix in dynamic programming). This implementation is on higher level of abstraction, so other representations can be used (artificial neural networks, tree etc.). For the testing of Q — Learning algorithm, maze solving problem was visualized in MATLAB programming language with the found trajectory marked on the maze. The maze in this paper is defined with starting position in the top left corner and the exit in the bottom right corner. The performance of the algorithm is measured for different scales of the problem.