Reinforcement learning (RL) constitutes an effective method of controlling dynamic systems without prior knowledge. One of the most important and difficult problems in RL is the improvement of data efficiency. Probabilistic inference for learning control (PILCO) is a state-of-the-art data-efficient framework that uses a Gaussian process to model dynamic systems. However, it only focuses on optimizing cumulative rewards and does not consider the accuracy of a dynamic model, which is an important factor for controller learning. To further improve the data efficiency of PILCO, we propose its active exploration version (AEPILCO) that utilizes information entropy to describe samples. In the policy evaluation stage, we incorporate an information entropy criterion into long-term sample prediction. Through the informative policy evaluation function, our algorithm obtains informative policy parameters in the policy improvement stage. Using the policy parameters in the actual execution produces an informative sample set; this is helpful in learning an accurate dynamic model. Thus, the AEPILCO algorithm improves data efficiency by learning an accurate dynamic model by actively selecting informative samples based on the information entropy criterion. We demonstrate the validity and efficiency of the proposed algorithm for several challenging controller problems involving a cart pole, a pendubot, a double pendulum, and a cart double pendulum. The AEPILCO algorithm can learn a controller using fewer trials compared to PILCO. This is verified through theoretical analysis and experimental results.
Depeng Gao, Jiafeng Liu, Rui Wu, Dansong Cheng, Xiaopeng Fan and Xianglong Tang
With the advent of 3D cameras, getting depth information along with RGB images has been facilitated, which is helpful in various computer vision tasks. However, there are two challenges in using these RGB-D images to help recognize RGB images captured by conventional cameras: one is that the depth images are missing at the testing stage, the other is that the training and test data are drawn from different distributions as they are captured using different equipment. To jointly address the two challenges, we propose an asymmetrical transfer learning framework, wherein three classifiers are trained using the RGB and depth images in the source domain and RGB images in the target domain with a structural risk minimization criterion and regularization theory. A cross-modality co-regularizer is used to restrict the two-source classifier in a consistent manner to increase accuracy. Moreover, an L2,1 norm cross-domain co-regularizer is used to magnify significant visual features and inhibit insignificant ones in the weight vectors of the two RGB classifiers. Thus, using the cross-modality and cross-domain co-regularizer, the knowledge of RGB-D images in the source domain is transferred to the target domain to improve the target classifier. The results of the experiment show that the proposed method is one of the most effective ones.