Explanation-Based Learning For Mobile-Robot Perception

摘要

Abstract Results in robot learning have demonstrated that robots can learn simple strategies from very little initial knowledge when in restricted environments [9, 2, 17]. While these results indicate the potential role of machine learning for robot perception and control, new approaches are needed to scale up to more complex problems and realistic environments. The fundamental roadblock to scaling up learning algorithms is that, as the complexity of the learning task (for example, the number of distinct sensors, complexity of individual sensor inputs, amount of sensor noise, complexity of the function to be learned) increases, it becomes increasingly difficult to correctly generalize from the limited quantity of training data available. Both experimental and theoretical research indicate that purely inductive learning methods, which basically rely on detecting statistical regularities among the training data, scale poorly with such increases in complexity [7, 30, 10].