Artificial Intelligence最新文献

Compared to autonomous agent learning, lifelong agent learning tackles the additional challenge of accumulating skills in a way favourable to long term development. What an agent learns at a given moment can be an element for the future creation of behaviours of greater complexity, whose purpose cannot be anticipated.

Beyond its initial low-level sensorimotor development phase, the agent is expected to acquire, in the same manner as skills, values and goals which support the development of complex behaviours beyond the reactive level. To do so, it must have a way to represent and memorize such information.

In this article, we identify the properties suitable for a representation system supporting the lifelong development of agents through a review of a wide range of memory systems and related literature. Following this analysis, our second contribution is the proposition and implementation of such a representation system in MIND, a modular architecture for lifelong development. The new variable module acts as a simple memory system which is strongly integrated to the hierarchies of skill modules of MIND, and allows for the progressive structuration of behaviour around persistent non-symbolic representations. Variable modules have many applications for the development and structuration of complex behaviours, but also offer designers and operators explicit models of values and goals facilitating human interaction, control and explainability.

We show through experiments two possible uses of variable modules. In the first experiment, skills exchange information by using a variable representing the concept of “target”, which allows the generalization of navigation behaviours. In the second experiment, we show how a non-symbolic representation can be learned and memorized to develop beyond simple reactive behaviour, and keep track of the steps of a process whose state cannot be inferred by observing the environment.

The subgraph isomorphism problem (SIP) is a challenging problem with wide practical applications. In the last decade, despite being a theoretical hard problem, researchers designed various algorithms for solving SIP. In this work, we propose five main strategies and develop an improved exact algorithm for SIP. First, we design a probing search procedure to try whether the search procedure can successfully obtain a solution at first sight. Second, we design a novel matching ordering strategy as a value-ordering heuristic, which uses some useful information obtained from the probing search procedure to preferentially select some promising target vertices. Third, we discuss the characteristics of different propagation methods in the context of SIP and present an adaptive propagation method to make a good balance between these methods. Moreover, to further improve the performance of solving large graphs, we propose an enhanced implementation of the edge constraint method and a domain limitation strategy, which aims to accelerate the search process. Experimental results on a broad range of classic and graph-database benchmarks show that our proposed algorithm performs better than several state-of-the-art algorithms for the SIP.

Counterfactual Explanations (CEs) have emerged as a major paradigm in explainable AI research, providing recourse recommendations for users affected by the decisions of machine learning models. However, CEs found by existing methods often become invalid when slight changes occur in the parameters of the model they were generated for. The literature lacks a way to provide exhaustive robustness guarantees for CEs under model changes, in that existing methods to improve CEs' robustness are mostly heuristic, and the robustness performances are evaluated empirically using only a limited number of retrained models. To bridge this gap, we propose a novel interval abstraction technique for parametric machine learning models, which allows us to obtain provable robustness guarantees for CEs under a possibly infinite set of plausible model changes Δ. Based on this idea, we formalise a robustness notion for CEs, which we call Δ-robustness, in both binary and multi-class classification settings. We present procedures to verify Δ-robustness based on Mixed Integer Linear Programming, using which we further propose algorithms to generate CEs that are Δ-robust. In an extensive empirical study involving neural networks and logistic regression models, we demonstrate the practical applicability of our approach. We discuss two strategies for determining the appropriate hyperparameters in our method, and we quantitatively benchmark CEs generated by eleven methods, highlighting the effectiveness of our algorithms in finding robust CEs.

MaxSAT is the problem of finding an assignment satisfying the maximum number of clauses in a CNF formula. We consider a natural generalization of this problem to generic sets of polynomials and propose a weighted version of Polynomial Calculus to address this problem.

Weighted Polynomial Calculus is a natural generalization of the systems MaxSAT-Resolution and weighted Resolution. Unlike such systems, weighted Polynomial Calculus manipulates polynomials with coefficients in a finite field and either weights in $N$ or $Z$. We show the soundness and completeness of weighted Polynomial Calculus via an algorithmic procedure.

Weighted Polynomial Calculus, with weights in $N$ and coefficients in $F_2$, is able to prove efficiently that Tseitin formulas on a connected graph are minimally unsatisfiable. Using weights in $Z$, it also proves efficiently that the Pigeonhole Principle is minimally unsatisfiable.

In this article, we present a novel approximation approach for abstract argumentation using a customized Graph Convolutional Network (GCN) architecture and a tailored training method. Our approach demonstrates promising results in approximating abstract argumentation tasks across various semantics, setting a new state of the art for performance on certain tasks. We provide a detailed analysis of approximation and runtime performance and propose a new scheme for evaluation. By advancing the state of the art for approximating the acceptability status of abstract arguments, we make theoretical and empirical advances in understanding the limits and opportunities for approximation in this field. Our approach shows potential for creating both general purpose and task-specific approximators and offers insights into the performance differences across benchmarks and semantics.

Surrogate modelling techniques have seen growing attention in recent years when applied to both modelling and optimisation of industrial design problems. These techniques are highly relevant when assessing the performance of a particular design carries a high cost, as the overall cost can be mitigated via the construction of a model to be queried in lieu of the available high-cost source. The construction of these models can sometimes employ other sources of information which are both cheaper and less accurate. The existence of these sources however poses the question of which sources should be used when constructing a model. Recent studies have attempted to characterise harmful data sources to guide practitioners in choosing when to ignore a certain source. These studies have done so in a synthetic setting, characterising sources using a large amount of data that is not available in practice. Some of these studies have also been shown to potentially suffer from bias in the benchmarks used in the analysis. In this study, we approach the characterisation of harmful low-fidelity sources as an algorithm selection problem. We employ recently developed benchmark filtering techniques to conduct a bias-free assessment, providing objectively varied benchmark suites of different sizes for future research. Analysing one of these benchmark suites with the technique known as Instance Space Analysis, we provide an intuitive visualisation of when a low-fidelity source should be used. By performing this analysis using only the limited data available to train a surrogate model, we are able to provide guidelines that can be directly used in an applied industrial setting.

We investigate an open question in the study of the curse of dimensionality: Is it possible to find the single nearest neighbor of a query in high dimensions? Using the notion of (in)distinguishability to examine whether the feature map of a kernel is able to distinguish two distinct points in high dimensions, we analyze this ability of a metric-based Lipschitz continuous kernel as well as that of the recently introduced Isolation Kernel. Between the two kernels, we show that only Isolation Kernel has distinguishability and it performs consistently well in four tasks: indexed search for exact nearest neighbor search, anomaly detection using kernel density estimation, t-SNE visualization and SVM classification in both low and high dimensions, compared with distance, Gaussian and three other existing kernels.

Dealing with controversial information is an important issue in several application contexts. Formal argumentation enables reasoning on arguments for and against a claim to decide on an outcome. Dung's abstract Argumentation Framework (AF) has emerged as a central formalism in argument-based reasoning. Key aspects of the success and popularity of Dung's framework include its simplicity and expressiveness. Integrity constraints help to express domain knowledge in a compact and natural way, thus keeping easy the modeling task even for problems that otherwise would be hard to encode within an AF. In this paper, we first explore two intuitive semantics based on Kleene and Lukasiewicz logics, respectively, for AF augmented with (strong) constraints—the resulting argumentation framework is called Constrained AF (CAF). Then, we propose a new argumentation framework called Weak constrained AF (WAF) that enhances CAF with weak constraints. Intuitively, these constraints can be used to find “optimal” solutions to problems defined through CAF. We provide a detailed complexity analysis of CAF and WAF, showing that strong constraints do not increase the expressive power of AF in most cases, while weak constraints systematically increase the expressive power of CAF (and AF) under several well-known argumentation semantics.

The structured argumentation system that represents arguments by premise-conclusion pairs is called premise-conclusion argumentation (PA) and the one that represents arguments by their premises is called base argumentation (BA). To assess whether BA and PA have the same ability in argument evaluation by extensional semantics, this paper defines the notion of extensional equivalence between BA and PA. It also defines the notion of bisimulation between BA and PA and shows that bisimulation implies extensional equivalence. To illustrate how base argumentation, bisimulation and extensional equivalence can contribute to the study of PA, we prove some new results about PA by investigating the extensional properties of a base argumentation framework and exporting them to two premise-conclusion argumentation frameworks via bisimulation and extensional equivalence. We show that there are essentially three kinds of extensions in these frameworks and that the extensions in the two premise-conclusion argumentation frameworks are identical.

We present a conceptualization providing an original domain-independent perspective on two crucial properties in reasoning: consistency and reinstatement. They emerge as a pair of dual characteristics, representing complementary requirements on the outcomes of reasoning processes. Central to our formalization are two underlying parametric relations: incompatibility and reinstatement violation. Different instances of these relations give rise to a spectrum of consistency and reinstatement scenarios. As a demonstration of versatility and expressive power of our approach we provide a characterization of various abstract argumentation semantics which are expressed as combinations of distinct consistency and reinstatement constraints. Moreover, we conduct an investigation into preserving these essential properties across different reasoning stages. Specifically, we delve into scenarios where a labelling is derived from other labellings through a synthesis function, using the synthesis of argument justification as an illustrative instance. We achieve a general characterization of consistency preservation synthesis functions, while we unveil an impossibility result concerning reinstatement preservation, leading us to explore an alternative notion to ensure feasibility. Our exploration reveals a weakness in the traditional definition of argument justification, for which we propose a refined version overcoming this limitation.