Feature learning for multi-task inverse reinforcement learning

In this paper we study the question of life long learning of behaviors from human demonstrations by an intelligent system. One approach is to model the observed demonstrations by a stationary policy. Inverse rein-forcement learning, on the other hand, searches a reward function that makes the observed policy closed to optimal in the corresponding Markov decision process. This approach provides a model of the task solved by the demonstrator and has been shown to lead to better generalization in un-known contexts. However both approaches focus on learning a single task from the expert demonstration. In this paper we propose a feature learn-ing approach for inverse reinforcement learning in which several different tasks are demonstrated, but in which each task is modeled as a mixture of several, simpler, primitive tasks. We present an algorithm based on an al-ternate gradient descent to learn simultaneously a dictionary of primitive tasks (in the form of reward functions) and their combination into an ap-proximation of the task underlying observed behavior. We illustrate how this approach enables efficient re-use of knowledge from previous demon-strations. Namely knowledge on tasks that were previously observed by the learner is used to improve the learning of a new composite behavior, thus achieving transfer of knowledge between tasks