Browsing by Author "Hsueh, Bo-Yang"

Now showing 1 - 2 of 2

Combining Deep Deterministic Policy Gradient with Cross-Entropy Method
(2019-01-01)
Lai, Tung-Yi
;
Hsueh, Chu-Hsuan
;
Lin, You-Hsuan
;
Chu, Yeong-Jia Roger
;
Hsueh, Bo-Yang
This paper proposes a deep reinforcement learning algorithm for solving robotic tasks, such as grasping objects. We propose in this paper a combination of cross-entropy optimization (CE) with deep deterministic policy gradient (DDPG). More specifically, where in the CE method, we first sample from a Gaussian distribution with zero as its initial mean, we now set the initial mean to DDPG's output instead. The resulting algorithm is referred to as the DDPG-CE method. Next, to negate the effects of bad samples, we improve on DDPG-CE by substituting the CE component with a weighted CE method, resulting in the DDPG-WCE algorithm. Experiments show that DDPG-WCE achieves a higher success rate on grasping previously unseen objects, than other approaches, such as supervised learning, DDPG, CE, and DDPG-CE.
Stochastic Gradient Descent with Hyperbolic-Tangent Decay on Classification
(2019-01-01)
Hsueh, Bo-Yang
;
Li, Wei
;
Wu, I-Chen
;
資訊工程學系
;
Department of Computer Science
Learning rate scheduler has been a critical issue in the deep neural network training. Several schedulers and methods have been proposed, including step decay scheduler, adaptive method, cosine scheduler and cyclical scheduler. This paper proposes a new scheduling method, named hyperbolic-tangent decay (HTD). We run experiments on several benchmarks such as: ResNet, Wide ResNet and DenseNet for CIFAR-10 and CIFAR-100 datasets, LSTM for PAMAP2 dataset, ResNet on ImageNet and Fashion-MNIST datasets. In our experiments, HTD outperforms step decay and cosine scheduler in nearly all cases, while requiring less hyperparameters than step decay, and more flexible than cosine scheduler. Code is available at https: //github.com/BIGBALLON/HTD.