Multi-task reinforcement learning endeavors to accomplish a set of different tasks with a single policy. To enhance data efficiency by sharing parameters across multiple tasks, a common practice segments the network into distinct modules and trains a routing network to recombine these modules into task-specific policies. However, existing routing approaches employ a fixed number of modules for all tasks, neglecting that tasks with varying difficulties commonly require varying amounts of knowledge. This work presents a Dynamic Depth Routing (D2R) framework, which learns strategic skipping of certain intermediate modules, thereby flexibly choosing different numbers of modules for each task. Under this framework, we further introduce a ResRouting method to address the issue of disparate routing paths between behavior and target policies during off-policy training. In addition, we design an automatic route-balancing mechanism to encourage continued routing exploration for unmastered tasks without disturbing the routing of mastered ones. We conduct extensive experiments on various robotics manipulation tasks in the Meta-World benchmark, where D2R achieves state-of-the-art performance with significantly improved learning efficiency.
相關內容
Existing off-policy reinforcement learning algorithms typically necessitate an explicit state-action-value function representation, which becomes problematic in high-dimensional action spaces. These algorithms often encounter challenges where they struggle with the curse of dimensionality, as maintaining a state-action-value function in such spaces becomes data-inefficient. In this work, we propose a novel off-policy trust region optimization approach, called Vlearn, that eliminates the requirement for an explicit state-action-value function. Instead, we demonstrate how to efficiently leverage just a state-value function as the critic, thus overcoming several limitations of existing methods. By doing so, Vlearn addresses the computational challenges posed by high-dimensional action spaces. Furthermore, Vlearn introduces an efficient approach to address the challenges associated with pure state-value function learning in the off-policy setting. This approach not only simplifies the implementation of off-policy policy gradient algorithms but also leads to consistent and robust performance across various benchmark tasks. Specifically, by removing the need for a state-action-value function Vlearn simplifies the learning process and allows for more efficient exploration and exploitation in complex environments
Amidst task-specific learning-based control synthesis frameworks that achieve impressive empirical results, a unified framework that systematically constructs an optimal policy for sufficiently solving a general notion of a task is absent. Hence, we propose a theoretical framework for a task-centered control synthesis leveraging two critical ideas: 1) oracle-guided policy optimization for the non-limiting integration of sub-optimal task-based priors to guide the policy optimization and 2) task-vital multimodality to break down solving a task into executing a sequence of behavioral modes. The proposed approach results in highly agile parkour and diving on a 16-DoF dynamic bipedal robot. The obtained policy advances indefinitely on a track, performing leaps and jumps of varying lengths and heights for the parkour task. Corresponding to the dive task, the policy demonstrates front, back, and side flips from various initial heights. Finally, we introduce a novel latent mode space reachability analysis to study our policies' versatility and generalization by computing a feasible mode set function through which we certify a set of failure-free modes for our policy to perform at any given state.
Graphic designers often get inspiration through the recombination of references. Our formative study (N=6) reveals that graphic designers focus on conceptual keywords during this process, and want support for discovering the keywords, expanding them, and exploring diverse recombination options of them, while still having room for designers' creativity. We propose CreativeConnect, a system with generative AI pipelines that helps users discover useful elements from the reference image using keywords, recommends relevant keywords, generates diverse recombination options with user-selected keywords, and shows recombinations as sketches with text descriptions. Our user study (N=16) showed that CreativeConnect helped users discover keywords from the reference and generate multiple ideas based on them, ultimately helping users produce more design ideas with higher self-reported creativity compared to the baseline system without generative pipelines. While CreativeConnect was shown effective in ideation, we discussed how CreativeConnect can be extended to support other types of tasks in creativity support.
Fine-grained control over large language models (LLMs) remains a significant challenge, hindering their adaptability to diverse user needs. While Reinforcement Learning from Human Feedback (RLHF) shows promise in aligning LLMs, its reliance on scalar rewards often limits its ability to capture diverse user preferences in real-world applications. To address this limitation, we introduce the Directional Preference Alignment (DPA) framework. Unlike the scalar-reward RLHF, DPA incorporates multi-objective reward modeling to represent diverse preference profiles. Additionally, DPA models user preferences as directions (i.e., unit vectors) in the reward space to achieve user-dependent preference control. Our method involves training a multi-objective reward model and then fine-tuning the LLM with a preference-conditioned variant of Rejection Sampling Finetuning (RSF), an RLHF method adopted by Llama 2. This method enjoys a better performance trade-off across various reward objectives. In comparison with the scalar-reward RLHF, DPA offers users intuitive control over LLM generation: they can arithmetically specify their desired trade-offs (e.g., more helpfulness with less verbosity). We also validate the effectiveness of DPA with real-world alignment experiments on Mistral-7B. Our method provides straightforward arithmetic control over the trade-off between helpfulness and verbosity while maintaining competitive performance with strong baselines such as Direct Preference Optimization (DPO).
We present SeaEval, a benchmark for multilingual foundation models. In addition to characterizing how these models understand and reason with natural language, we also investigate how well they comprehend cultural practices, nuances, and values. Alongside standard accuracy metrics, we investigate the brittleness of foundation models in the dimensions of semantics and multilinguality. Our analyses span both open-sourced and closed models, leading to empirical results across classic NLP tasks, reasoning, and cultural comprehension. Key findings indicate (1) Most models exhibit varied behavior when given paraphrased instructions. (2) Many models still suffer from exposure bias (e.g., positional bias, majority label bias). (3) For questions rooted in factual, scientific, and commonsense knowledge, consistent responses are expected across multilingual queries that are semantically equivalent. Yet, most models surprisingly demonstrate inconsistent performance on these queries. (4) Multilingually-trained models have not attained "balanced multilingual" capabilities. Our endeavors underscore the need for more generalizable semantic representations and enhanced multilingual contextualization. SeaEval can serve as a launchpad for more thorough investigations and evaluations for multilingual and multicultural scenarios.
A key feature of federated learning (FL) is to preserve the data privacy of end users. However, there still exist potential privacy leakage in exchanging gradients under FL. As a result, recent research often explores the differential privacy (DP) approaches to add noises to the computing results to address privacy concerns with low overheads, which however degrade the model performance. In this paper, we strike the balance of data privacy and efficiency by utilizing the pervasive social connections between users. Specifically, we propose SCFL, a novel Social-aware Clustered Federated Learning scheme, where mutually trusted individuals can freely form a social cluster and aggregate their raw model updates (e.g., gradients) inside each cluster before uploading to the cloud for global aggregation. By mixing model updates in a social group, adversaries can only eavesdrop the social-layer combined results, but not the privacy of individuals. We unfold the design of SCFL in three steps.i) Stable social cluster formation. Considering users' heterogeneous training samples and data distributions, we formulate the optimal social cluster formation problem as a federation game and devise a fair revenue allocation mechanism to resist free-riders. ii) Differentiated trust-privacy mapping}. For the clusters with low mutual trust, we design a customizable privacy preservation mechanism to adaptively sanitize participants' model updates depending on social trust degrees. iii) Distributed convergence}. A distributed two-sided matching algorithm is devised to attain an optimized disjoint partition with Nash-stable convergence. Experiments on Facebook network and MNIST/CIFAR-10 datasets validate that our SCFL can effectively enhance learning utility, improve user payoff, and enforce customizable privacy protection.
The existence of representative datasets is a prerequisite of many successful artificial intelligence and machine learning models. However, the subsequent application of these models often involves scenarios that are inadequately represented in the data used for training. The reasons for this are manifold and range from time and cost constraints to ethical considerations. As a consequence, the reliable use of these models, especially in safety-critical applications, is a huge challenge. Leveraging additional, already existing sources of knowledge is key to overcome the limitations of purely data-driven approaches, and eventually to increase the generalization capability of these models. Furthermore, predictions that conform with knowledge are crucial for making trustworthy and safe decisions even in underrepresented scenarios. This work provides an overview of existing techniques and methods in the literature that combine data-based models with existing knowledge. The identified approaches are structured according to the categories integration, extraction and conformity. Special attention is given to applications in the field of autonomous driving.
Semi-supervised learning on class-imbalanced data, although a realistic problem, has been under studied. While existing semi-supervised learning (SSL) methods are known to perform poorly on minority classes, we find that they still generate high precision pseudo-labels on minority classes. By exploiting this property, in this work, we propose Class-Rebalancing Self-Training (CReST), a simple yet effective framework to improve existing SSL methods on class-imbalanced data. CReST iteratively retrains a baseline SSL model with a labeled set expanded by adding pseudo-labeled samples from an unlabeled set, where pseudo-labeled samples from minority classes are selected more frequently according to an estimated class distribution. We also propose a progressive distribution alignment to adaptively adjust the rebalancing strength dubbed CReST+. We show that CReST and CReST+ improve state-of-the-art SSL algorithms on various class-imbalanced datasets and consistently outperform other popular rebalancing methods.
Meta reinforcement learning (meta-RL) extracts knowledge from previous tasks and achieves fast adaptation to new tasks. Despite recent progress, efficient exploration in meta-RL remains a key challenge in sparse-reward tasks, as it requires quickly finding informative task-relevant experiences in both meta-training and adaptation. To address this challenge, we explicitly model an exploration policy learning problem for meta-RL, which is separated from exploitation policy learning, and introduce a novel empowerment-driven exploration objective, which aims to maximize information gain for task identification. We derive a corresponding intrinsic reward and develop a new off-policy meta-RL framework, which efficiently learns separate context-aware exploration and exploitation policies by sharing the knowledge of task inference. Experimental evaluation shows that our meta-RL method significantly outperforms state-of-the-art baselines on various sparse-reward MuJoCo locomotion tasks and more complex sparse-reward Meta-World tasks.
State-of-the-art Convolutional Neural Network (CNN) benefits a lot from multi-task learning (MTL), which learns multiple related tasks simultaneously to obtain shared or mutually related representations for different tasks. The most widely-used MTL CNN structure is based on an empirical or heuristic split on a specific layer (e.g., the last convolutional layer) to minimize different task-specific losses. However, this heuristic sharing/splitting strategy may be harmful to the final performance of one or multiple tasks. In this paper, we propose a novel CNN structure for MTL, which enables automatic feature fusing at every layer. Specifically, we first concatenate features from different tasks according to their channel dimension, and then formulate the feature fusing problem as discriminative dimensionality reduction. We show that this discriminative dimensionality reduction can be done by 1x1 Convolution, Batch Normalization, and Weight Decay in one CNN, which we refer to as Neural Discriminative Dimensionality Reduction (NDDR). We perform ablation analysis in details for different configurations in training the network. The experiments carried out on different network structures and different task sets demonstrate the promising performance and desirable generalizability of our proposed method.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191