Maximum Entropy Deep Inverse Reinforcement Learning

This paper presents a general framework for employing deep architectures – in particular neural networks – to solve the inverse reinforcement learning (IRL) problem. Specifically, we propose to exploit the representational capacity and favourable computational complexity of deep networks to approximate complex, nonlinear reward functions in scenarios with large state spaces. We show that the Maximum Entropy paradigm for IRL lends itself naturally to the efficient training of deep architectures.At test time, the approach leads to a computational complexity independent of the number of demonstrations. This makes it especially well-suited for applications in life-long learning scenarios commonly encountered in robotics.We demonstrate that our approach achieves performance commensurate to the state-of-the-art on existing benchmarks already with simple, comparatively shallow network architectures while significantly outperforming the state-of-the-art on an alternative benchmark based on more…


Link to Full Article: Maximum Entropy Deep Inverse Reinforcement Learning