Real-world robots commonly have to act in complex, poorly understood environments where the true world dynamics are unknown. To compensate for the unknown world dynamics, we often provide a class of models to a learner so it may select a model, typically using a minimum prediction error metric over a set of training data. Often in real-world domains the model class is unable to capture the true dynamics, due to either limited domain knowledge or a desire to use a small model. In these cases we call the model class misspecified, and an unfortunate consequence of misspecification is that even with unlimited data and computation there is no guarantee the model with minimum prediction error leads to the best performing policy. In this work, our approach improves upon the standard maximum likelihood model selection metric by explicitly selecting the model which achieves the highest expected reward, rather than the most likely model. We present an algorithm for which the highest performing model from the model class is guaranteed to be found given unlimited data and computation. Empirically, we demonstrate that our algorithm is often superior to the maximum likelihood learner in a batch learning setting for two common RL benchmark problems and a third real-world system, the hydrodynamic cart-pole, a domain whose complex dynamics cannot be known exactly.

Reinforcement Learning offers a very general framework for learning controllers, but its effectiveness is closely tied to the controller parameterization used. Especially when learning feedback controllers for weakly stable systems, ineffective parameterizations can result in unstable controllers and poor performance both in terms of learning convergence and in the cost of the resulting policy. In this paper we explore four linear controller parameterizations in the context of REINFORCE, applying them to the control of a reaching task with a linearized flexible manipulator. We find that some natural but naive parameterizations perform very poorly, while the Youla Parameterization (a popular parameterization from the controls literature) offers a number of robustness and performance advantages.

Policy gradient (PG) reinforcement learning algorithms have strong (local) convergence guarantees, but their learning performance is typically limited by a large variance in the estimate of the gradient. In this paper, we formulate the variance reduction problem by describing a signal-to-noise ratio (SNR) for policy gradient algorithms, and evaluate this SNR carefully for the popular Weight Perturbation (WP) algorithm. We confirm that SNR is a good predictor of long-term learning performance, and that in our episodic formulation, the cost-to-go function is indeed the optimal baseline. We then propose two modifications to traditional model-free policy gradient algorithms in order to optimize the SNR. First, we examine WP using anisotropic sampling distributions, which introduces a bias into the update but increases the SNR; this bias can be interpretted as following the natural gradient of the cost function. Second, we show that non-Gaussian distributions can also increase the SNR, and argue that the optimal isotropic distribution is a ‘shell’ distribution with a constant magnitude and uniform distribution in direction. We demonstrate that both modifications produce substantial improvements in learning performance in challenging policy gradient experiments.