IJCAI 2023 论文列表

Recently, graph neural networks (GNNs) have drawn a surge of investigations in deep graph clustering. Nevertheless, existing approaches predominantly are inclined to semantic-agnostic since GNNs exhibit inherent limitations in capturing global underlying semantic structures. Meanwhile, multiple objectives are imposed within one latent space, whereas representations from different granularities may presumably conflict with each other, yielding severe performance degradation for clustering. To this end, we propose a novel Multi-Level Graph Contrastive Prototypical Clustering (MLG-CPC) framework for end-to-end clustering. Specifically, a Prototype Discrimination (ProDisc) objective function is proposed to explicitly capture semantic information via cluster assignments. Moreover, to alleviate the issue of objectives conflict, we introduce to perceive representations of different granularities within individual feature-, prototypical-, and cluster-level spaces by the feature decorrelation, prototype contrast, and cluster space consistency respectively. Extensive experiments on four benchmarks demonstrate the superiority of the proposed MLG-CPC against the state-of-the-art graph clustering approaches.

The robustness of graph classification models plays an essential role in providing highly reliable applications. Previous studies along this line primarily focus on seeking the stability of the model in terms of overall data metrics (e.g., accuracy) when facing data perturbations, such as removing edges. Empirically, we find that these graph classification models also suffer from semantic bias and confidence collapse issues, which substantially hinder their applicability in real-world scenarios. To address these issues, we present MGRL, a multi-view representation learning model for graph classification tasks that achieves robust results. Firstly, we proposes an instance-view consistency representation learning method, which utilizes multi-granularity contrastive learning technique to perform semantic constraints on instance representations at both the node and graph levels, thus alleviating the semantic bias issue. Secondly, we proposes a class-view discriminative representation learning method, which employs the prototype-driven class distance optimization technique to adjust intra- and inter-class distances, thereby mitigating the confidence collapse issue.Finally, extensive experiments and visualizations on eight benchmark dataset demonstrate the effectiveness of MGRL.

Multi-Task Learning (MTL) is a well-established paradigm for learning shared models for a diverse set of tasks. Moreover, MTL improves data efficiency by jointly training all tasks simultaneously. However, directly optimizing the losses of all the tasks may lead to imbalanced performance on all the tasks due to the competition among tasks for the shared parameters in MTL models. Many MTL methods try to mitigate this problem by dynamically weighting task losses or manipulating task gradients. Different from existing studies, in this paper, we propose a Neural Ordinal diffeRential equation based Multi-tAsk Learning (NORMAL) method to alleviate this issue by modeling task-specific feature transformations from the perspective of dynamic flows built on the Neural Ordinary Differential Equation (NODE). Specifically, the proposed NORMAL model designs a time-aware neural ODE block to learn task-specific time information, which determines task positions of feature transformations in the dynamic flow, in NODE automatically via gradient descent methods. In this way, the proposed NORMAL model handles the problem of competing shared parameters by learning task positions. Moreover, the learned task positions can be used to measure the relevance among different tasks. Extensive experiments show that the proposed NORMAL model outperforms state-of-the-art MTL models.

Graph-level contrastive learning, aiming to learn the representations for each graph by contrasting two augmented graphs, has attracted considerable attention. Previous studies usually simply assume that a graph and its augmented graph as a positive pair, otherwise as a negative pair. However, it is well known that graph structure is always complex and multi-scale, which gives rise to a fundamental question: after graph augmentation, will the previous assumption still hold in reality? By an experimental analysis, we discover the semantic information of an augmented graph structure may be not consistent as original graph structure, and whether two augmented graphs are positive or negative pairs is highly related with the multi-scale structures. Based on this finding, we propose a multi-scale subgraph contrastive learning architecture which is able to characterize the fine-grained semantic information. Specifically, we generate global and local views at different scales based on subgraph sampling, and construct multiple contrastive relationships according to their semantic associations to provide richer self-supervised signals. Extensive experiments and parametric analyzes on eight graph classification real-world datasets well demonstrate the effectiveness of the proposed method.

With the rapid progress in Multi-Agent Path Finding (MAPF), researchers have studied how MAPF algorithms can be deployed to coordinate hundreds of robots in large automated warehouses. While most works try to improve the throughput of such warehouses by developing better MAPF algorithms, we focus on improving the throughput by optimizing the warehouse layout. We show that, even with state-of-the-art MAPF algorithms, commonly used human-designed layouts can lead to congestion for warehouses with large numbers of robots and thus have limited scalability. We extend existing automatic scenario generation methods to optimize warehouse layouts. Results show that our optimized warehouse layouts (1) reduce traffic congestion and thus improve throughput, (2) improve the scalability of the automated warehouses by doubling the number of robots in some cases, and (3) are capable of generating layouts with user-specified diversity measures.

In recent years, many convolutional neural network-based models are designed for JPEG artifacts reduction, and have achieved notable progress. However, few methods are suitable for extreme low-bitrate image compression artifacts reduction. The main challenge is that the highly compressed image loses too much information, resulting in reconstructing high-quality image difficultly. To address this issue, we propose a multimodal fusion learning method for text-guided JPEG artifacts reduction, in which the corresponding text description not only provides the potential prior information of the highly compressed image, but also serves as supplementary information to assist in image deblocking. We fuse image features and text semantic features from the global and local perspectives respectively, and design a contrastive loss built upon contrastive learning to produce visually pleasing results. Extensive experiments, including a user study, prove that our method can obtain better deblocking results compared to the state-of-the-art methods.

In the study of reactive systems, qualitative properties are usually easier to model and analyze than quantitative properties. This is especially true in systems where mutually beneficial cooperation between agents is possible, such as multi-agent systems. The large number of possible payoffs available to agents in reactive systems with quantitative properties means that there are many scenarios in which agents deviate from mutually beneficial outcomes in order to gain negligible payoff improvements. This behavior often leads to less desirable outcomes for all agents involved. For this reason we study satisficing goals, derived from a decision-making approach aimed at meeting a good-enough outcome instead of pure optimization. By considering satisficing goals, we are able to employ efficient automata-based algorithms to find pure-strategy Nash equilibria. We then show that these algorithms extend to scenarios in which agents have multiple thresholds, providing an approximation of optimization while still retaining the possibility of mutually beneficial cooperation and efficient automata-based algorithms. Finally, we demonstrate a one-way correspondence between the existence of epsilon-equilibria and the existence of equilibria in games where agents have multiple thresholds.

For an agent in a multi-agent environment, it is often beneficial to be able to predict what other agents will do next when deciding how to act. Previous work in multi-agent intention scheduling assumes a priori knowledge of the current goals of other agents. In this paper, we present a new approach to multi-agent intention scheduling in which an agent uses online goal recognition to identify the goals currently being pursued by other agents while acting in pursuit of its own goals. We show how online goal recognition can be incorporated into an MCTS-based intention scheduler, and evaluate our approach in a range of scenarios. The results demonstrate that our approach can rapidly recognise the goals of other agents even when they are pursuing multiple goals concurrently, and has similar performance to agents which know the goals of other agents a priori.

In the last decade, there have been significant advances in multi-agent reinforcement learning (MARL) but there are still numerous challenges, such as high sample complexity and slow convergence to stable policies, that need to be overcome before wide-spread deployment is possible. However, many real-world environments already, in practice, deploy sub-optimal or heuristic approaches for generating policies. An interesting question that arises is how to best use such approaches as advisors to help improve reinforcement learning in multi-agent domains. We provide a principled framework for incorporating action recommendations from online sub-optimal advisors in multi-agent settings. We describe the problem of ADvising Multiple Intelligent Reinforcement Agents (ADMIRAL) in nonrestrictive general-sum stochastic game environments and present two novel Q-learning-based algorithms: ADMIRAL - Decision Making (ADMIRAL-DM) and ADMIRAL - Advisor Evaluation (ADMIRAL-AE), which allow us to improve learning by appropriately incorporating advice from an advisor (ADMIRAL-DM), and evaluate the effectiveness of an advisor (ADMIRAL-AE). We analyze the algorithms theoretically and provide fixed point guarantees regarding their learning in general-sum stochastic games. Furthermore, extensive experiments illustrate that these algorithms: can be used in a variety of environments, have performances that compare favourably to other related baselines, can scale to large state-action spaces, and are robust to poor advice from advisors.

As autonomous systems tackle more real-world situations, mission success oftentimes cannot be guaranteed and the planner must reason about the probability of failure. Unfortunately, computing a trajectory that satisfies mission goals while constraining the probability of failure is difficult because of the need to reason about complex, multidimensional probability distributions. Recent methods have seen success using chance-constrained, model-based planning. We argue there are two main drawbacks to these approaches. First, current methods suffer from an inability to deal with expressive environment models such as 3D non-convex obstacles. Second, most planners rely on considerable simplifications when computing trajectory risk including approximating the agent's dynamics, geometry, and uncertainty. We apply hybrid search to the risk-bound, goal-directed planning problem. The hybrid search consists of a region planner and a trajectory planner. The region planner makes discrete choices by reasoning about geometric regions that the agent should visit in order to accomplish its mission. In formulating the region planner, we propose landmark regions that help produce obstacle-free paths. The region planner passes paths through the environment to a trajectory planner; the task of the trajectory planner is to optimize trajectories that respect the agent's dynamics and the user's desired risk of mission failure. We discuss three approaches to modeling trajectory risk: a CDF-based approach, a sampling-based collocation method, and an algorithm named Shooting Method Monte Carlo. A variety of 2D and 3D test cases are presented in the full paper including a linear case, a Dubins car model, and an underwater autonomous vehicle. The method is shown to outperform other methods in terms of speed and utility of the solution. Additionally, the models of trajectory risk are shown to better approximate risk in simulation.

In an offline reinforcement learning setting, the safe policy improvement (SPI) problem aims to improve the performance of a behavior policy according to which sample data has been generated. State-of-the-art approaches to SPI require a high number of samples to provide practical probabilistic guarantees on the improved policy's performance. We present a novel approach to the SPI problem that provides the means to require less data for such guarantees. Specifically, to prove the correctness of these guarantees, we devise implicit transformations on the data set and the underlying environment model that serve as theoretical foundations to derive tighter improvement bounds for SPI. Our empirical evaluation, using the well-established SPI with baseline bootstrapping (SPIBB) algorithm, on standard benchmarks shows that our method indeed significantly reduces the sample complexity of the SPIBB algorithm.

A moral planning agent (MPA) seeks to compare two plans or compute an optimal plan in an interactive setting with other agents, where relative ideality and optimality of plans are defined with respect to a prioritized value base. We model MPAs whose values are expressed by formulas of linear temporal logic (LTL) and define comparison for both joint plans and individual plans. We introduce different evaluation criteria for individual plans including an optimistic (risk-seeking) criterion, a pessimistic (risk-averse) one, and two criteria based on the use of anticipated responsibility. We provide complexity results for a variety of MPA problems.

Molecular de novo design is a critical yet challenging task in scientific fields, aiming to design novel molecular structures with desired property profiles. Significant progress has been made by resorting to generative models for graphs. However, limited attention is paid to hierarchical generative models, which can exploit the inherent hierarchical structure (with rich semantic information) of the molecular graphs and generate complex molecules of larger size that we shall demonstrate to be difficult for most existing models. The primary challenge to hierarchical generation is the non-differentiable issue caused by the generation of intermediate discrete coarsened graph structures. To sidestep this issue, we cast the tricky hierarchical generation problem over discrete spaces as the reverse process of hierarchical representation learning and propose MolHF, a new hierarchical flow-based model that generates molecular graphs in a coarse-to-fine manner. Specifically, MolHF first generates bonds through a multi-scale architecture, then generates atoms based on the coarsened graph structure at each scale. We demonstrate that MolHF achieves state-of-the-art performance in random generation and property optimization, implying its high capacity to model data distribution. Furthermore, MolHF is the first flow-based model that can be applied to model larger molecules (polymer) with more than 100 heavy atoms. The code and models are available at https://github.com/violet-sto/MolHF.

Training deep learning models on data distributed and local to edge devices such as mobile phones is a prominent recent research direction. In a Gossip Learning (GL) system, each participating device maintains a model trained on its local data and iteratively aggregates it with the models from its neighbours in a communication network. While the fully distributed operation in GL comes with natural advantages over the centralized orchestration in Federated Learning (FL), its convergence becomes particularly slow when the data distribution is heterogeneous and aligns with the clustered structure of the communication network. These characteristics are pervasive across practical applications as people with similar interests (thus producing similar data) tend to create communities. This paper proposes a data-driven neighbor weighting strategy for aggregating the models: this enables faster diffusion of knowledge across the communities in the network and leads to quicker convergence. We augment the method to make it computationally efficient and fair: the devices quickly converge to the same model. We evaluate our model on real and synthetic datasets that we generate using a novel generative model for communication networks with heterogeneous data. Our exhaustive empirical evaluation verifies that our proposed method attains a faster convergence rate than the baselines. For example, the median test accuracy for a decentralized bird image classifier application reaches 81% with our proposed method within 80 rounds, whereas the baseline only reaches 46%.

There is an increasing demand to design policy interventions to achieve various targets specified by the UN Sustainable Development Goals by 2030. Designing interventions is a complex task given that the system may often respond in unexpected ways to a given intervention. This could be due to interventions towards a given target, affecting other unrelated variables, and/or interventions leading to acute disparities in nearby geographic areas. In order to address such issues, we propose a novel concept called Stress Modeling that analyzes the holistic impact of a policy intervention by taking into account the interactions within a system, after the intervention. The simulation is based on the postulate that complex systems of interacting entities tend to settle down into "low energy'' configurations by minimizing differentials in capabilities of neighbouring entities. The simulation shows how policy impact percolates through geospatial boundaries over time and can be applied at any granularity. The theory and the corresponding package have been explained along with a case study analyzing a fertilizer policy in the Agro-climatic Zones of the state of Karnataka, India.

Practical uses of Artificial Intelligence (AI) in the real world have demonstrated the importance of embedding moral choices into intelligent agents. They have also highlighted that defining top-down ethical constraints on AI according to any one type of morality is extremely challenging and can pose risks. A bottom-up learning approach may be more appropriate for studying and developing ethical behavior in AI agents. In particular, we believe that an interesting and insightful starting point is the analysis of emergent behavior of Reinforcement Learning (RL) agents that act according to a predefined set of moral rewards in social dilemmas. In this work, we present a systematic analysis of the choices made by intrinsically-motivated RL agents whose rewards are based on moral theories. We aim to design reward structures that are simplified yet representative of a set of key ethical systems. Therefore, we first define moral reward functions that distinguish between consequence- and norm-based agents, between morality based on societal norms or internal virtues, and between single- and mixed-virtue (e.g., multi-objective) methodologies. Then, we evaluate our approach by modeling repeated dyadic interactions between learning moral agents in three iterated social dilemma games (Prisoner's Dilemma, Volunteer's Dilemma and Stag Hunt). We analyze the impact of different types of morality on the emergence of cooperation, defection or exploitation, and the corresponding social outcomes. Finally, we discuss the implications of these findings for the development of moral agents in artificial and mixed human-AI societies.

In this paper we investigate the optimal controller synthesis problem, so that the system under the controller can reach a specified target set while satisfying given constraints. Existing model predictive control (MPC) methods learn from a set of discrete states visited by previous (sub-)optimized trajectories and thus result in computationally expensive mixed-integer nonlinear optimization. In this paper a novel MPC method is proposed based on reach-avoid analysis to solve the controller synthesis problem iteratively. The reach-avoid analysis is concerned with computing a reach-avoid set which is a set of initial states such that the system can reach the target set successfully. It not only provides terminal constraints, which ensure feasibility of MPC, but also expands discrete states in existing methods into a continuous set (i.e., reach-avoid sets) and thus leads to nonlinear optimization which is more computationally tractable online due to the absence of integer variables. Finally, we evaluate the proposed method and make comparisons with state-of-the-art ones based on several examples.

While massive valuable deep models trained on large-scale data have been released to facilitate the artificial intelligence community, they may encounter attacks in deployment which leads to privacy leakage of training data. In this work, we propose a learning approach termed differentially private data-free distillation (DPDFD) for model conversion that can convert a pretrained model (teacher) into its privacy-preserving counterpart (student) via an intermediate generator without access to training data. The learning collaborates three parties in a unified way. First, massive synthetic data are generated with the generator. Then, they are fed into the teacher and student to compute differentially private gradients by normalizing the gradients and adding noise before performing descent. Finally, the student is updated with these differentially private gradients and the generator is updated by taking the student as a fixed discriminator in an alternate manner. In addition to a privacy-preserving student, the generator can generate synthetic data in a differentially private way for other down-stream tasks. We theoretically prove that our approach can guarantee differential privacy and well convergence. Extensive experiments that significantly outperform other differentially private generative approaches demonstrate the effectiveness of our approach.

We study the Improving Multi-Armed Bandit problem, where the reward obtained from an arm increases with the number of pulls it receives. This model provides an elegant abstraction for many real-world problems in domains such as education and employment, where decisions about the distribution of opportunities can affect the future capabilities of communities and the disparity between them. A decision-maker in such settings must consider the impact of her decisions on future rewards in addition to the standard objective of maximizing her cumulative reward at any time. We study the tension between two seemingly conflicting objectives in the horizon-unaware setting: a) maximizing the cumulative reward at any time and b) ensuring that arms with better long-term rewards get sufficient pulls even if they initially have low rewards. We show that, surprisingly, the two objectives are aligned with each other. Our main contribution is an anytime algorithm for the IMAB problem that achieves the best possible cumulative reward while ensuring that the arms reach their true potential given sufficient time. Our algorithm mitigates the initial disparity due to lack of opportunity and continues pulling an arm until it stops improving. We prove the optimality of our algorithm by showing that a) any algorithm for the IMAB problem, no matter how utilitarian, must suffer Omega(T) policy regret and Omega(k) competitive ratio with respect to the optimal offline policy, and b) the competitive ratio of our algorithm is O(k).