Broad Learning from Narrow Training: A Case Study in Robotic Soccer

The range of unseen instances that can be successfully classiied by a learning algorithm is determined not only by the distribution of the training data, but also by the parameters of the function to be learned. With the right parameters, learning in a certain region of the state space can generalize to completely diierent areas with no retraining. We demonstrate the power of using well-chosen inputs to (and outputs from) neural networks by conducting experiments in our robotic soccer domain. We train an agent to shoot a moving ball into a goal in a speciic situation and end up with a general shooting behavior that is much more widely applicable.