Deep inverse planning for learning from high-dimensional demonstrations
In the context of planning problems, observing human demonstrations allows a system to efficiently learn a policy (mapping of states to actions) through methods such as inverse optimal control (IOC). However, many complex problems are embedded within decision processes which require exceedingly high-dimensional representations. Unfortunately, traditional approaches to IOC become intractable when faced with this high-dimensionality. Recent work in deep reinforcement learning addresses this limitation by learning neural network parameters for estimating the expected value of actions from a given state. Once trained, the network efficiently maps raw visual input to a model free policy. However, this process fails when feedback from the environment is not available and the reward/cost function requires a non-linear high-dimensional parameterization. We propose extending recent work in deep reinforcement learning to the inverse planning setting for efficiently learning complex behaviors from high-dimensional data without the need for environment feedback (rewards, coordinates, etc.) or low-dimensional reward/cost functions.
