Artificial Intelligence最新文献

This paper presents a formal model of dialectical argument strength in terms of the number of ways in which an argument can be successfully attacked in expansions of an abstract argumentation framework. First a model is proposed that is abstract but designed to avoid overly limiting assumptions on instantiations or dialogue contexts. It is then shown that most principles for argument strength proposed in the literature fail to hold for the proposed notions of dialectical strength, which clarifies the rational foundations of these principles and highlights the importance of distinguishing between kinds of argument strength, in particular logical, dialectical and rhetorical argument strength. The abstract model is then instantiated with ASPIC⁺ to test the claim that it does not make overly limiting assumptions on the structure of arguments and the nature of their relations.

Meta-learning has arisen as a powerful tool for many machine learning problems. With multiple factors to be considered when designing learning models for real-world applications, meta-learning with multiple objectives has attracted much attention recently. However, existing works either linearly combine multiple objectives into one objective or adopt evolutionary algorithms to handle it, where the former approach needs to pay high computational cost to tune the combination coefficients while the latter approach is computationally heavy and incapable to be integrated into gradient-based optimization. To alleviate those limitations, in this paper, we aim to propose a generic gradient-based Multi-Objective Meta-Learning (MOML) framework with applications in many machine learning problems. Specifically, the MOML framework formulates the objective function of meta-learning with multiple objectives as a Multi-Objective Bi-Level optimization Problem (MOBLP) where the upper-level subproblem is to solve several possibly conflicting objectives for the meta-learner. Different from those existing works, in this paper, we propose a gradient-based algorithm to solve the MOBLP. Specifically, we devise the first gradient-based optimization algorithm by alternately solving the lower-level and upper-level subproblems via the gradient descent method and the gradient-based multi-objective optimization method, respectively. Theoretically, we prove the convergence property and provide a non-asymptotic analysis of the proposed gradient-based optimization algorithm. Empirically, extensive experiments justify our theoretical results and demonstrate the superiority of the proposed MOML framework for different learning problems, including few-shot learning, domain adaptation, multi-task learning, neural architecture search, and reinforcement learning. The source code of MOML is available at https://github.com/Baijiong-Lin/MOML.

Although the development of AI in games is remarkable, intelligent machines still lag behind humans in games that require the ability of language understanding. In this paper, we focus on the crossword puzzle resolution task. Solving crossword puzzles is a challenging task since it requires the ability to answer natural language questions with knowledge and the ability to execute a search over possible answers to find an optimal set of solutions for the grid. Previous solutions are devoted to exploiting heuristic strategies in search to find solutions while having limited ability to explore the search space. We build a comprehensive system for crossword puzzle resolution based on Monte Carlo Tree Search (MCTS). As far as we know, we are the first to model the crossword puzzle resolution problem as a Markov Decision Process and apply the MCTS to solve it. We construct a dataset for crossword puzzle resolution based on daily puzzles from The New York Times with detailed specifications of both the puzzle and clue database selection. Our method achieves state-of-the-art performance on the dataset. The code of the system and experiments in this paper is publicly available: https://www.github.com/lhlclhl/CP.

As reinforcement learning methods increasingly amass accomplishments, the need for comprehending their solutions becomes more crucial. Most explainable reinforcement learning (XRL) methods generate a static explanation depicting their developers' intuition of what should be explained and how. In contrast, literature from the social sciences proposes that meaningful explanations are structured as a dialog between the explainer and the explainee, suggesting a more active role for the user and her communication with the agent. In this paper, we present ASQ-IT – an interactive explanation system that presents video clips of the agent acting in its environment based on queries given by the user that describe temporal properties of behaviors of interest. Our approach is based on formal methods: queries in ASQ-IT's user interface map to a fragment of Linear Temporal Logic over finite traces (LTLf), which we developed, and our algorithm for query processing is based on automata theory. User studies show that end-users can understand and formulate queries in ASQ-IT and that using ASQ-IT assists users in identifying faulty agent behaviors.

Human-aware planning involves generating plans that are explicable, i.e. conform to user expectations, as well as providing explanations when such plans cannot be found. In this paper, we bring these two concepts together and show how an agent can achieve a trade-off between these two competing characteristics of a plan. To achieve this, we conceive a first-of-its-kind planner MEGA that can reason about the possibility of explaining a plan in the plan generation process itself. We will also explore how solutions to such problems can be expressed as “self-explaining plans” – and show how this representation allows us to leverage classical planning compilations of epistemic planning to reason about this trade-off at plan generation time without having to incur the computational burden of having to search in the space of differences between the agent model and the mental model of the human in the loop in order to come up with the optimal trade-off. We will illustrate these concepts in two well-known planning domains, as well as with a robot in a typical search and reconnaissance task. Human factor studies in the latter highlight the usefulness of the proposed approach.

Qualitative probabilistic networks (QPNs) combine the conditional independence assumptions of Bayesian networks with the qualitative properties of positive and negative dependence. They formalise various intuitive properties of positive dependence to allow inferences over a large network of variables. However, we will demonstrate in this paper that, due to an incorrect symmetry property, many inferences obtained in non-binary QPNs are not mathematically true. We will provide examples of such incorrect inferences and briefly discuss possible resolutions.

The evaluation of the fidelity of eXplainable Artificial Intelligence (XAI) methods to their underlying models is a challenging task, primarily due to the absence of a ground truth for explanations. However, assessing fidelity is a necessary step for ensuring a correct XAI methodology. In this study, we conduct a fair and objective comparison of the current state-of-the-art XAI methods by introducing three novel image datasets with reliable ground truth for explanations. The primary objective of this comparison is to identify methods with low fidelity and eliminate them from further research, thereby promoting the development of more trustworthy and effective XAI techniques. Our results demonstrate that XAI methods based on the direct gradient calculation and the backpropagation of output information to input yield higher accuracy and reliability compared to methods relying on perturbation based or Class Activation Maps (CAM). However, these methods tend to generate more noisy saliency maps. These findings have significant implications for the advancement of XAI methods, enabling the elimination of erroneous explanations and fostering the development of more robust and reliable XAI.

We initiate the study of fairness among classes of agents in online bipartite matching where there is a given set of offline vertices (aka agents) and another set of vertices (aka items) that arrive online and must be matched irrevocably upon arrival. In this setting, agents are partitioned into classes and the matching is required to be fair with respect to the classes. We adapt popular fairness notions (e.g. envy-freeness, proportionality, and maximin share) and their relaxations to this setting and study deterministic algorithms for matching indivisible items (leading to integral matchings) and for matching divisible items (leading to fractional matchings). For matching indivisible items, we propose an adaptive-priority-based algorithm, Match-and-Shift, prove that it achieves $\frac{1}{2}$-approximation of both class envy-freeness up to one item and class maximin share fairness, and show that each guarantee is tight. For matching divisible items, we design a water-filling-based algorithm, Equal-Filling, that achieves $(1-\frac{1}{e})$-approximation of class envy-freeness and class proportionality; we prove $1-\frac{1}{e}$ to be tight for class proportionality and establish a $\frac{3}{4}$ upper bound on class envy-freeness. Finally, we discuss several challenges in designing randomized algorithms that achieve reasonable fairness approximation ratios. Nonetheless, we build upon Equal-Filling to design a randomized algorithm for matching indivisible items, Equal-Filling-OCS, which achieves 0.593-approximation of class proportionality.

Structural entropy is a metric that measures the amount of information embedded in graph structure data under a strategy of hierarchical abstracting. To measure the structural entropy of a dynamic graph, we need to decode the optimal encoding tree corresponding to the best community partitioning for each snapshot. However, the current methods do not support dynamic encoding tree updating and incremental structural entropy computation. To address this issue, we propose Incre-2dSE, a novel incremental measurement framework that dynamically adjusts the community partitioning and efficiently computes the updated structural entropy for each updated graph. Specifically, Incre-2dSE includes incremental algorithms based on two dynamic adjustment strategies for two-dimensional encoding trees, i.e., the naive adjustment strategy and the node-shifting adjustment strategy, which support theoretical analysis of updated structural entropy and incrementally optimize community partitioning towards a lower structural entropy. We conduct extensive experiments on 3 artificial datasets generated by Hawkes Process and 3 real-world datasets. Experimental results confirm that our incremental algorithms effectively capture the dynamic evolution of the communities, reduce time consumption, and provide great interpretability.

Controlled Query Evaluation (CQE) is a framework for the protection of confidential data, where a policy given in terms of logic formulae indicates which information must be kept private. Functions called censors filter query answering so that no answers are returned that may lead a user to infer data protected by the policy. The preferred censors, called optimal censors, are the ones that conceal only what is necessary, thus maximizing the returned answers. Typically, given a policy over a data or knowledge base, several optimal censors exist.

Our research on CQE is based on the following intuition: confidential data are those that violate the logical assertions specifying the policy, and thus censoring them in query answering is similar to processing queries in the presence of inconsistent data as studied in Consistent Query Answering (CQA). In this paper, we investigate the relationship between CQE and CQA in the context of Description Logic ontologies. We borrow the idea from CQA that query answering is a form of skeptical reasoning that takes into account all possible optimal censors. This approach leads to a revised notion of CQE, which allows us to avoid making an arbitrary choice on the censor to be selected, as done by previous research on the topic.

We then study the data complexity of query answering in our CQE framework, for conjunctive queries issued over ontologies specified in the popular Description Logics ${\text{DL-Lite}}_R$ and ${\mathrm{EL}}_{\perp }$. In our analysis, we consider some variants of the censor language, which is the language used by the censor to enforce the policy. Whereas the problem is in general intractable for simple censor languages, we show that for ${\text{DL-Lite}}_R$ ontologies it is first-order rewritable, and thus in AC⁰ in data complexity, for the most expressive censor language we propose.