Due to its strong ability in distribution representation, the diffusion model has been incorporated into offline reinforcement learning (RL) to cover diverse trajectories of the complex behavior policy. However, this also causes several challenges. Training the diffusion model to imitate behavior from the collected trajectories suffers from limited stitching capability which derives better policies from suboptimal trajectories. Furthermore, the inherent randomness of the diffusion model can lead to unpredictable control and dangerous behavior for the robot. To address these concerns, we propose the value-learning-based decision diffuser (V-DD), which consists of the trajectory diffusion module (TDM) and the trajectory evaluation module (TEM). During the training process, the TDM combines the state-value and classifier-free guidance to bolster the ability to stitch suboptimal trajectories. During the inference process, we design the TEM to select a feasible trajectory generated by the diffusion model. Empirical results demonstrate that our method delivers competitive results on the D4RL benchmark and substantially outperforms current diffusion model-based methods on the real-world robot task.
由于扩散模型在分布表示方面的强大能力,它已被纳入离线强化学习(RL),以覆盖复杂行为政策的各种轨迹。然而,这也带来了一些挑战。从收集到的轨迹中训练扩散模型来模仿行为,会受到拼接能力的限制,从而从次优轨迹中得出更好的策略。此外,扩散模型固有的随机性可能会导致机器人无法预测的控制和危险行为。为了解决这些问题,我们提出了基于价值学习的决策扩散器(V-DD),它由轨迹扩散模块(TDM)和轨迹评估模块(TEM)组成。在训练过程中,TDM 结合了状态值和无分类器指导,以提高缝合次优轨迹的能力。在推理过程中,我们设计 TEM 来选择由扩散模型生成的可行轨迹。实证结果表明,我们的方法在 D4RL 基准测试中取得了具有竞争力的结果,并且在实际机器人任务中大大优于当前基于扩散模型的方法。
{"title":"Stabilizing Diffusion Model for Robotic Control With Dynamic Programming and Transition Feasibility","authors":"Haoran Li;Yaocheng Zhang;Haowei Wen;Yuanheng Zhu;Dongbin Zhao","doi":"10.1109/TAI.2024.3387401","DOIUrl":"https://doi.org/10.1109/TAI.2024.3387401","url":null,"abstract":"Due to its strong ability in distribution representation, the diffusion model has been incorporated into offline reinforcement learning (RL) to cover diverse trajectories of the complex behavior policy. However, this also causes several challenges. Training the diffusion model to imitate behavior from the collected trajectories suffers from limited stitching capability which derives better policies from suboptimal trajectories. Furthermore, the inherent randomness of the diffusion model can lead to unpredictable control and dangerous behavior for the robot. To address these concerns, we propose the value-learning-based decision diffuser (V-DD), which consists of the trajectory diffusion module (TDM) and the trajectory evaluation module (TEM). During the training process, the TDM combines the state-value and classifier-free guidance to bolster the ability to stitch suboptimal trajectories. During the inference process, we design the TEM to select a feasible trajectory generated by the diffusion model. Empirical results demonstrate that our method delivers competitive results on the D4RL benchmark and substantially outperforms current diffusion model-based methods on the real-world robot task.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-04-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142165059","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-04-10DOI: 10.1109/TAI.2024.3387406
Kangjia He;Li Liu;Youmin Zhang;Ye Wang;Qun Liu;Guoyin Wang
Counterfactual subgraphs explain graph neural networks (GNNs) by answering the question: “How would the prediction change if a certain subgraph were absent in the input instance?” The differentiable proxy adjacency matrix is prevalent in current counterfactual subgraph discovery studies due to its ability to avoid exhaustive edge searching. However, a prediction gap exists when feeding the proxy matrix with continuous values and the thresholded discrete adjacency matrix to GNNs, compromising the optimization of the subgraph generator. Furthermore, the end-to-end learning schema adopted in the subgraph generator limits the diversity of counterfactual subgraphs. To this end, we propose CF-GFNExplainer, a flow-based approach for learning counterfactual subgraphs. CF-GFNExplainer employs a policy network with a discrete edge removal schema to construct counterfactual subgraph generation trajectories. Additionally, we introduce a loss function designed to guide CF-GFNExplainer's optimization. The discrete adjacency matrix generated in each trajectory eliminates the prediction gap, enhancing the validity of the learned subgraphs. Furthermore, the multitrajectories sampling strategy adopted in CF-GFNExplainer results in diverse counterfactual subgraphs. Extensive experiments conducted on synthetic and real-world datasets demonstrate the effectiveness of the proposed method in terms of validity and diversity.
{"title":"Learning Counterfactual Explanation of Graph Neural Networks via Generative Flow Network","authors":"Kangjia He;Li Liu;Youmin Zhang;Ye Wang;Qun Liu;Guoyin Wang","doi":"10.1109/TAI.2024.3387406","DOIUrl":"https://doi.org/10.1109/TAI.2024.3387406","url":null,"abstract":"Counterfactual subgraphs explain graph neural networks (GNNs) by answering the question: “How would the prediction change if a certain subgraph were absent in the input instance?” The differentiable proxy adjacency matrix is prevalent in current counterfactual subgraph discovery studies due to its ability to avoid exhaustive edge searching. However, a prediction gap exists when feeding the proxy matrix with continuous values and the thresholded discrete adjacency matrix to GNNs, compromising the optimization of the subgraph generator. Furthermore, the end-to-end learning schema adopted in the subgraph generator limits the diversity of counterfactual subgraphs. To this end, we propose CF-GFNExplainer, a flow-based approach for learning counterfactual subgraphs. CF-GFNExplainer employs a policy network with a discrete edge removal schema to construct counterfactual subgraph generation trajectories. Additionally, we introduce a loss function designed to guide CF-GFNExplainer's optimization. The discrete adjacency matrix generated in each trajectory eliminates the prediction gap, enhancing the validity of the learned subgraphs. Furthermore, the multitrajectories sampling strategy adopted in CF-GFNExplainer results in diverse counterfactual subgraphs. Extensive experiments conducted on synthetic and real-world datasets demonstrate the effectiveness of the proposed method in terms of validity and diversity.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-04-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142169597","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-04-10DOI: 10.1109/TAI.2024.3386499
Ziyan Zhang;Bo Jiang;Jin Tang;Jinhui Tang;Bin Luo
Graph convolutional networks (GCNs) gain increasing attention on graph data learning tasks in recent years. However, in many applications, graph may come with an incomplete form where attributes of graph nodes are partially unknown/missing. Existing graph convolutions (GCs) are generally designed on complete graphs which cannot deal with attribute-incomplete graph data directly. To address this problem, in this article, we extend standard GC and develop an explicit Partial Graph Convolution (PaGC) for attribute-incomplete graph data. Our PaGC is derived based on the observation that the core neighborhood aggregator in GC operation can be equivalently viewed as an energy minimization model. Based on it, we can define a novel partial aggregation function