Deeply AggreVaTeD: Differentiable Imitation Learning for Sequential\n Prediction

摘要

Researchers have demonstrated state-of-the-art performance in sequential\ndecision making problems (e.g., robotics control, sequential prediction) with\ndeep neural network models. One often has access to near-optimal oracles that\nachieve good performance on the task during training. We demonstrate that\nAggreVaTeD --- a policy gradient extension of the Imitation Learning (IL)\napproach of (Ross & Bagnell, 2014) --- can leverage such an oracle to achieve\nfaster and better solutions with less training data than a less-informed\nReinforcement Learning (RL) technique. Using both feedforward and recurrent\nneural network predictors, we present stochastic gradient procedures on a\nsequential prediction task, dependency-parsing from raw image data, as well as\non various high dimensional robotics control problems. We also provide a\ncomprehensive theoretical study of IL that demonstrates we can expect up to\nexponentially lower sample complexity for learning with AggreVaTeD than with RL\nalgorithms, which backs our empirical findings. Our results and theory indicate\nthat the proposed approach can achieve superior performance with respect to the\noracle when the demonstrator is sub-optimal.\n