Planning how to learn

When a robot uses an imperfect system model to plan its actions, a key challenge is the exploration-exploitation trade-off between two sometimes conflicting objectives: (i) learning and improving the model, and (ii) immediate progress towards the goal, according to the current model. To address model uncertainty systematically, we propose to use Bayesian reinforcement learning and cast it as a partially observable Markov decision process (POMDP). We present a simple algorithm for offline POMDP planning in the continuous state space. Offline planning produces a POMDP policy, which can be executed efficiently online as a finite-state controller. This approach seamlessly integrates planning and learning: it incorporates learning objectives in the computed plan, which then enables the robot to learn nearly optimally online and reach the goal. We evaluated the approach in simulations on two distinct tasks, acrobot swing-up and autonomous vehicle navigation amidst pedestrians, and obtained interesting preliminary results.