The Introspective Robot: Using Self-Prediction to Improve Robot Learning

We investigate the use of self-predicting neural networks for autonomous robot learning within noisy or partially predictable environments. A benchmark experiment is performed in which a network is trained on a task consisting of a mixture of predictable and random patterns. In addition to learning the task patterns, the network is also trained to explicitly predict the internal representations developed for each pattern as well as the resulting output error. Selfprediction is found to speed up learning and may offer an effective framework for distinguishing predictable from unpredictable input data.