Using reinforcement learning to solve the labyrinth game, a nonlinear control application

Many control problems have been successfully solved using artificial neural nets, e.g the truck backer-upper problem and the pole-chart problem to name a few. A complex problem is the labyrinth game control application, where the goal is to guide a ball through a maze, avoiding obstacles on the way to a target position by altering the angles of the board accordingly. Thus, the labyrinth is a dynamic nonlinear control problem. One way to make a neural network learn this control task on-line, is the reinforcement learning strategy. This paper presents a solution to the labyrinth control problem based on a hybrid algorithm combining a SRV neuron and regular backpropagation neurons, together with results regarding the hardware application