A Robot That Improves Its Ability To Learn
Joseph O'Sullivan and Sebastian Thrun
The use of machine learning is attractive in the design of autonomous robots, since it enables robots to adapt to the unforeseen. However, one of the key bottlenecks of current machine learning algorithms is the enormous sample complexity, which appears to prohibit their usage in all but the simplest of robotic domains. To make machine learning more practical in such domains, more powerful algorithms are needed that can generalize more accurately from less training data.
This paper investigates the feasibility of learning algorithms that gradually improve over the lifetime of the robot. When faced with a novel thing to learn, knowledge acquired in previous learning tasks improves the ability of the robot to generalize, hence reduces the sample complexity. In this paper, we report results of applying a particular algorithm to mobile robot perception problems, which was originally proposed by Suddarth [Sud90]. The learning tasks considered involve the recognition of persons, objects and locations. We illustrate that having previously learned related tasks allows a robot to learn a novel task from significantly fewer training examples than if it was learning from scratch. Based on these results, we argue that self-adjusting learning strategies are superior to conventional learning algorithms in many robotic domains.