In the realm of complex environments for robot control tasks, Meta Reinforcement Learning has emerged as a pivotal component, leveraging prior experiences to tackle unseen, long-term, and sparsely rewarded intricate tasks. Skill-based Meta-RL methods aim to extract useful skills from task contexts, aiding agents in swiftly adapting to new environments during meta-testing. However, the skills learned by existing methods lack generality and adaptability, limiting their performance across meta-testing task sets. To address this, a Skill Discovery Meta-RL (SDMRL) approach is proposed in this paper, by learning more useful skills in the absence of reward functions. The objective of the SDMRL is formalized as maximizing an information-theoretic objective using maximum entropy policies, enabling agents to learn valuable skills and skill priors from unstructured data in an unsupervised manner. Experimental results in continuous control tasks such as Maze Navigation demonstrate the effectiveness of the SDMRL approach over previous meta reinforcement learning methods, and the learned skills proficiently address the long-term complex sparse reward tasks.

Cross-network node classification aims to transfer knowledge extracted from a source network with sufficient labeled nodes to predict labels for nodes in a target network. Existing methods mainly focus on associating two networks in a shared embedding space, where the distribution discrepancy of node embedding across two networks isminimized. However, the relationship across two networks remains underexplored. In this paper, we propose a method that learns cross-network relationship among nodes from different networks and performs information aggregation across networks based on our learned relationship. To achieve this, we introduce a labeled Fused Gromov-Wasserstein model based on optimal transport, which exploits feature, structure and label information to construct node association between two networks. Based on the cross-network relationship, we design a cross-network graph convolutional network to learn node embeddings. Experimental results on several benchmark datasets show the superiority of the proposed method over state-of-the-art methods.

Dialogue systems are an integral part of human-computer interaction, and their goal is to achieve a natural, intelligent, and smooth dialog experience. However, traditional dialog systems face challenges in understanding user intent, generating reasonable responses, and maintaining dialog coherence. The core reason is the mismatch between the supervised training of the dialog system strategy network and the final evaluation metrics, and the objective function fails to cover the critical metrics required by the system. Some recent research work has begun to explore reinforcement learning-oriented training; however, the existing work still suffers from the problems of relying on artificially defined templates and too many dialog rounds. In order to solve the above problems, this study employs a Transformer-based dialog strategy network to construct a word-level deep reinforcement learning training process, which improves the interactive performance of the dialog system by combining the multifaceted metrics of the dialog system to construct rewards. The experimental results show that relative to the traditional supervised training method, the policy network for deep reinforcement learning has an improvement in the dialog performance metrics BLEU and Comb, and maintains comparable performance in other metrics. The contribution of this study is to apply reinforcement learning techniques to the training process of policy networks for dialog systems, and its effectiveness and potential are verified through experiments.