1. Probabilistic Multi-Dimensional Classification
   • The authors propose a formal framework for probabilistic multi-dimensional classification, which decomposes learning an optimal multi-dimensional classifier into learning a set of smaller single-variable multi-class probabilistic classifiers and a directed acyclic graph. The framework addresses challenges such as inaccuracy, scalability, limited use to certain types of data, hardness of interpretation, and lack of probabilistic estimations.
2. Policy Regularization with Dataset Constraint for Offline Reinforcement Learning
   • This paper presents the Policy Regularization with Dataset Constraint (PRDC) method to address limitations of offline reinforcement learning (RL) related to the lack of exploration. PRDC promotes training stability, learning efficiency, and final performance of existing offline RL methods during online fine-tuning across various locomotion and navigation tasks.
3. Modular Continual Learning with Probabilistic Framework
   • The authors develop a modular continual learning framework called PICLE that accelerates search by using a probabilistic model to compute the fitness of each composition. They show that using PICLE can achieve different types of transfer while scaling to large search spaces and offers better performance than existing methods.
4. Ensemble-based Offline-to-Online Reinforcement Learning: From Pessimistic Learning to Optimistic Exploration
   • The Ensemble-based Offline-to-Online (E2O) RL framework presented in this paper increases training stability, learning efficiency, and final performance of existing offline RL methods during online fine-tuning across a range of locomotion and navigation tasks.
5. Adversarial Constrained Bidding via Minimax Regret Optimization with Causality-Aware Reinforcement Learning
   • This paper presents a Minimax Regret Optimization (MiRO) approach that can be applied to the problem of constrained bidding in adversarial bidding environments. By incorporating expert demonstrations for learning bidding strategies, the MiRO with Causality-aware reinforcement Learning (MiROCL) algorithm outperforms prior methods by over 30%.
6. iPLAN: Intent-Aware Planning in Heterogeneous Traffic via Distributed Multi-Agent Reinforcement Learning
   • This paper presents iPLAN, a distributed multi-agent reinforcement learning (MARL) algorithm for intent-aware planning in dense and heterogeneous traffic scenarios. It demonstrates superior performance compared to centralized MARL baselines in both mild and chaotic traffic. The method achieves a higher episodic reward, higher success rate, and longer survival time than conventional approaches.
7. Ada-NAV: Adaptive Trajectory-Based Sample Efficient Policy Learning for Robotic Navigation
   • The authors propose Ada-NAV, an adaptive trajectory length scheme for reinforcement learning-based robotic navigation. This method emphasizes exploration at the beginning of the training and exploitation later on and demonstrates superior performance and sample efficiency compared to other approaches. It performs well in gridworld settings, simulated robotic environments, and real-world robotic experiments.
8. AROID: Improving Adversarial Robustness through Online Instance-wise Data Augmentation
   • This work presents AROID, a method that improves adversarial robustness for deep neural networks by automatically learning online, instance-wise data augmentation policies. The proposed approach effectively mitigates the ASR error propagation issue and outperforms existing methods on various synthetic and real-world datasets.
9. Multimodal Audio-textual Architecture for Robust Spoken Language Understanding
   • The paper investigates the impacts of ASR error propagation on state-of-the-art natural language understanding systems and proposes a multimodal language understanding module that utilizes self-supervised features learned from both audio and text modalities. The proposed approach is shown to be robust towards poor-quality ASR transcripts and outperforms conventional methods on various robotic manipulation and control tasks.
10. On the Efficacy of 3D Point Cloud Reinforcement Learning
    • This work explores the efficacy of 3D point cloud reinforcement learning by comparing 2D image and 3D point cloud RL methods on both minimalist synthetic tasks and complex robotic manipulation tasks. The results indicate that 3D point cloud RL can significantly outperform 2D methods when agent-object or object-object relationship encoding is a key factor.