下面是摘自IEEE机器人学习技术委员会(IEEE Technical Committe On Robot Learning)的一段话:

Learning techniques are increasingly being used in todays’ complex robotic system. Robots are expected to deal with a large variety of tasks, using their high-dimensional and complex bodies, to interact with objects and humans in an intuitive and friendly way. In this new setting, not all relevant information is available at design time, thus self-experimentation and learning by interacting with the physical and social world is very important to acquire knowledge.

A major obstacle, in high and complex sensorimotor space, is that learning can become extremely slow or even impossible without adequate exploration strategies. To solve this problem, two main approaches are now converging. Active learning, from statistical learning theory, where the learner actively chooses experiments in order to collect highly informative examples, and where expected information gain can be evaluated with either theoretically optimal criteria or various computationally efficient heuristics. The second approach, intrinsically motivated exploration, from developmental psychology and recently operationalized in the developmental robotics community, aims at building robots capable of open-ended cumulative learning through task-independent efficient exploration of their sensorimotor space and to refine our understanding of how children learn and develop.

Although similar in some aspects, these two approaches differ in some of the underlying assumptions. Active learning implicitly assumes that samples with high uncertainty are the most informative and focuses on single tasks. On the contrary, Intrinsic motivation has been identified by psychologists as an innate incentive that pushes organisms to spontaneously explore activities or situations for the sole reason that they have a certain degree of novelty, challenge or surprise, hence the term curiosity-driven learning sometimes used.

Advertisements