An adaptive ´broom balancer´ with visual inputs

An adaptive network with visual inputs has been trained to balance an inverted pendulum. Simulation results show that the network is capable of extracting the necessary state information from time sequences of crude visual images. A single linear adaptive threshold element (ADALINE) was adequate for this task. When tested by simulation, the performance achieved was sufficient to keep the pendulum from falling. The adaptive network´s ability to generalize made this possible since the training set encompassed only a fraction of all possible states.<>