Deep Gaussian Covariance Network with Trajectory Sampling for Data-Efficient Policy Search

Probabilistic world models increase data efficiency of model-based reinforcement learning (MBRL) by guiding the policy with their epistemic uncertainty to improve exploration and acquire new samples. Moreover, the uncertainty-aware learning procedures in probabilistic approaches lead to robust policies that are less sensitive to noisy observations compared to uncertainty unaware solutions. We propose to combine trajectory sampling and deep Gaussian covariance network (DGCN) for a data-efficient solution to MBRL problems in an optimal control setting. We compare trajectory sampling with density-based approximation for uncertainty propagation using three different probabilistic world models; Gaussian processes, Bayesian neural networks, and DGCNs. We provide empirical evidence using four different well-known test environments, that our method improves the sample-efficiency over other combinations of uncertainty propagation methods and probabilistic models. During our tests, we place particular emphasis on the robustness of the learned policies with respect to noisy initial states.

Gradient and Uncertainty Enhanced Sequential Sampling for Global Fit

Surrogate models based on machine learning methods have become an important part of modern engineering to replace costly computer simulations. The data used for creating a surrogate model are essential for the model accuracy and often restricted due to cost and time constraints. Adaptive sampling strategies have been shown to reduce the number of samples needed to create an accurate model. This paper proposes a new sampling strategy for global fit called Gradient and Uncertainty Enhanced Sequential Sampling (GUESS). The acquisition function uses two terms: the predictive posterior uncertainty of the surrogate model for exploration of unseen regions and a weighted approximation of the second and higher-order Taylor expansion values for exploitation. Although various sampling strategies have been proposed so far, the selection of a suitable method is not trivial. Therefore, we compared our proposed strategy to 9 adaptive sampling strategies for global surrogate modeling, based on 26 different 1 to 8-dimensional deterministic benchmarks functions. Results show that GUESS achieved on average the highest sample efficiency compared to other surrogate-based strategies on the tested examples. An ablation study considering the behavior of GUESS in higher dimensions and the importance of surrogate choice is also presented.

Deep Gaussian Covariance Network

The correlation length-scale next to the noise variance are the most used hyperparameters for the Gaussian processes. Typically, stationary covariance functions are used, which are only dependent on the distances between input points and thus invariant to the translations in the input space. The optimization of the hyperparameters is commonly done by maximizing the log marginal likelihood. This works quite well, if the distances are uniform distributed. In the case of a locally adapted or even sparse input space, the prediction of a test point can be worse dependent of its position. A possible solution to this, is the usage of a non-stationary covariance function, where the hyperparameters are calculated by a deep neural network. So that the correlation length scales and possibly the noise variance are dependent on the test point. Furthermore, different types of covariance functions are trained simultaneously, so that the Gaussian process prediction is an additive overlay of different covariance matrices. The right covariance functions combination and its hyperparameters are learned by the deep neural network. Additional, the Gaussian process will be able to be trained by batches or online and so it can handle arbitrarily large data sets. We call this framework Deep Gaussian Covariance Network (DGCP). There are also further extensions to this framework possible, for example sequentially dependent problems like time series or the local mixture of experts. The basic framework and some extension possibilities will be presented in this work. Moreover, a comparison to some recent state of the art surrogate model methods will be performed, also for a time dependent problem.