The range of unseen instances that can be successfully classified 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. That is to say, any learning algorithm is representation-dependent: with inappropriate representations, learning algorithms will not reach their full potential, but with well-chosen representations they can be quite powerful.
In particular, Neural Networks (NNs) have been successfully used to handle problems in motion control from pole balancing [12] to driving autonomous vehicles [7]. After training, they can generalize to previously unseen situations. However, the true extent of their flexibility is limited by the input representation. If their input is specific to a given situation in any way, then they will not generalize beyond that situation. For example, Pomerleau describes a situation in which his autonomous vehicle can learn to turn left, but then cannot turn right [7]. Therefore it is the job of the user to provide, if possible, inputs to a NN (or other learning algorithm) that are maximally flexible.
In this paper we illustrate and demonstrate the power of flexible inputs for the task of learning to shoot a moving ball into the goal in a robotic soccer domain: one team member passes the ball in front of another, which in turn must time its acceleration and adjust its heading so that it redirects the ball into the goal. This learning task is an example of a low-level behavior in robotic soccer: it is a complex individual task that is prerequisite for higher-level team-oriented strategy issues. We train a player to score in a narrow situation and then test how broadly applicable the learned behavior is. We use a NN to train the shooting teammate to accomplish this task, but we do so in such a way that the learning applies to a much broader range of situations than are covered by the initial training. The secret of this flexibility is our choice of inputs to the NN.
This paper is organized as follows. Section 2 situates this work within the growing body of robotic soccer research and briefly describes our simulator. Section 3 describes our experimental setup. Sections 4-7 contain details of the experiments along with results. Section 8 provides some discussion and conclusions. The contributions of this paper include creating a flexible representation for learning behaviors in a real-time control domain, such as robotic soccer; learning a particular complex behavior with a robust NN; and giving extensive empirical validation supporting the effectiveness of the representation used.