Machine Teaching of Active Sequential Learners

Machine Teaching is the problem of finding the best training data that can guide a learning algorithm to a target model with minimal effort. Teachers that provide data consistent with the true distribution for sequential learners who actively choose queries can be sub-optimal for finite horizons. A Markov decision process is used, the dynamics being a model of the learner and the actions being the teacher's responses. The paper also looks at the problem of learning from a teacher that plans.

Learning is improved when we plan teaching and the learner has a model of the teacher. Can be of use to model strategic planning behaviour of users of interactive intelligent systems by assuming they are boundedly optimal teachers.