Annals of Mathematics and Artificial Intelligence最新文献

In this paper we consider a new class of RBF (Radial Basis Function) neural networks, in which smoothing factors are replaced with shifts. We prove under certain conditions on the activation function that these networks are capable of approximating any continuous multivariate function on any compact subset of the d-dimensional Euclidean space. For RBF networks with finitely many fixed centroids we describe conditions guaranteeing approximation with arbitrary precision.

In this note, I present my personal view on the interaction of the three areas Automated Programming, Symbolic Computation, and Machine Learning. Programming is the activity of finding a (hopefully) correct program (algorithm) for a given problem. Programming is central to automation in all areas and is considered one of the most creative human activities. However, already very early in the history of programming, people started to “jump to the meta-level” of programming, i.e., started to develop procedures that automate, or semi-automate, (various aspects or parts of) the process of programming. This area has various names like “Automated Programming”, “Automated Algorithm Synthesis”, etc. Developing compilers can be considered an early example of a problem in automated programming. Automated reasoners for proving the correctness of programs with respect to a specification is an advanced example of a topic in automated programming. ChatGPT producing (amazingly good) programs from problem specifications in natural language is a recent example of automated programming. Programming tends to become the most important activity as the level of technological sophistication increases. Therefore, automating programming is maybe the most exciting and relevant technological endeavor today. It also will have enormous impact on the global job market in the software industry. Roughly, I see two main approaches to automated programming:

• symbolic computation

• and machine learning.

In this note, I explain how the two approaches work and that they are fundamentally different because they address two completely different ways of how problems are specified. Together, the two approaches constitute (part of) what some people like to call “artificial intelligence”. In my analysis, both approaches are just part of (algorithmic) mathematics. The approaches, like all non-trivial mathematical methods, need quite some intelligence on the side of the human inventors of the methods. However, applying the methods is just “machine execution” of algorithms. It is misleading to call the application “machine intelligence” or “artificial intelligence”. The analysis of the two approaches to automated programming also suggests that the two approaches, in the future, should be combined to achieve even higher levels of sophistication. At the end of this note, I propose some research questions for this new direction.

Agent Programming Languages have been studied for over 20 years for programming complex decision-making for autonomous systems. The GOAL agent programming language is particularly interesting since it depends on automated planning based on beliefs and goals to determine behavior rather than preprogrammed planning by developers. Model checking is a powerful verification technique to guarantee the safety of an autonomous system. Despite studies of model checking in other agent programming languages, GOAL lacks support for model checking of GOAL programs. The fundamental challenge is to make GOAL programs feasible for model checking. In this paper, we tackle this fundamental issue. First, we formalize the syntax and semantics of the logic underpinning stratified single-agent GOAL programs. Second, we devise an algorithm for transforming a stratified single-agent GOAL program to a transition system that is equivalent in terms of operational semantics, enabling model checking. Third, we develop an automated translator for a stratified single-agent GOAL program. The translator consists of (1) the automated transformation of a GOAL program into its operational semantically equivalent transition system, and (2) the interface generation of the generated transition system into a Prism model, an input for two probabilistic symbolic model checkers: Storm and Prism. Moreover, we point out that we will extend the applicability of the transformation algorithm and its implementation to all stratified GOAL programs.

The continuous Hopfield network (CHN) has provided a powerful approach to optimization problems and has shown good performance in different domains. However, two primary challenges still remain for this network: defining appropriate parameters and hyperparameters. In this study, our objective is to address these challenges and achieve optimal solutions for combinatorial optimization problems, thereby improving the overall performance of the continuous Hopfield network. To accomplish this, we propose a new technique for tuning the parameters of the CHN by considering its stability. To evaluate our approach, three well-known combinatorial optimization problems, namely, weighted constraint satisfaction problems, task assignment problems, and the traveling salesman problem, were employed. The experiments demonstrate that the proposed approach offers several advantages for CHN parameter tuning and the selection of optimal hyperparameter combinations.

Finding a collision-free feasible path for mobile robots is very important because they are essential in many fields such as healthcare, military, and industry. In this paper, a novel Clustering Obstacles (CO)-based path planning algorithm for mobile robots is presented using a quintic trigonometric Bézier curve and its two shape parameters. The CO-based algorithm forms clusters of geometrically shaped obstacles and finds the cluster centers. Moreover, the proposed waypoint algorithm (WP) finds the waypoints of the predefined skeleton path in addition to the start and destination points in an environment. Based on all these points, the predefined quintic trigonometric Bézier path candidates, taking the skeleton path as their convex hull, are then generated using the shape parameters of this curve. Moreover, the performance of the proposed algorithm is compared with K-Means and agglomerative hierarchical algorithms to obtain the quintic trigonometric Bézier paths desired by the user. The experimental results show that the CO-based path planning algorithm achieves better cluster centers and consequently better collision-free predefined paths.

The k-nearest neighbors (k/NN) algorithm is a simple yet powerful non-parametric classifier that is robust to noisy data and easy to implement. However, with the growing literature on k/NN methods, it is increasingly challenging for new researchers and practitioners to navigate the field. This review paper aims to provide a comprehensive overview of the latest developments in the k/NN algorithm, including its strengths and weaknesses, applications, benchmarks, and available software with corresponding publications and citation analysis. The review also discusses the potential of k/NN in various data science tasks, such as anomaly detection, dimensionality reduction and missing value imputation. By offering an in-depth analysis of k/NN, this paper serves as a valuable resource for researchers and practitioners to make informed decisions and identify the best k/NN implementation for a given application.

A digitized rigid motion is called digitally continuous if two neighbor pixels still stay neighbors after the motion. This concept plays important role when people or computers (artificial intelligence, machine vision) need to recognize the object shown in the image. In this paper, digital rotations of a pixel with its closest neighbors are of our interest. We compare the neighborhood motion map results among the three regular grids, when the center of rotation is the midpoint of a main pixel, a grid point (corner of a pixel) or an edge midpoint. The first measure about the quality of digital rotations is based on bijectivity, e.g., measuring how many of the cases produce bijective and how many produce not bijective neighborhood motion maps (Avkan et. al, 2022). Now, a second measure is investigated, the quality of bijective digital rotations is measured by the digital continuity of the resulted image: we measure how many of the cases are bijective and also digitally continuous. We show that rotations on the triangular grid prove to be digitally continuous at many more real angles and also as a special case, many more integer angles compared to the square grid or to the hexagonal grid with respect to the three different rotation centers.

A digital plane is a digitization of a Euclidean plane. A plane is specified by its normal, which is a 3D vector with integer coordinates, as considered in this case. It is established here that a 3D digital straight line segment, shifted by an integer amount, can produce the digitized plane. 3D plane’s normals are classified based on the Greatest Common Divisor (GCD) of its components, and the net code is calculated separately for each case. Experimental results are provided for several normals. Also, we show that the digital plane segment generated is a connected digital plane. The proposed method mainly involves integer arithmetic.

Hedonic games are a prominent model of coalition formation, in which each agent’s utility only depends on the coalition she resides. The subclass of hedonic games that models the formation of general partnerships (Larson 2018), where all affiliates receive the same utility, is referred to as hedonic games with common ranking property (HGCRP). Aside from their economic motivation, HGCRP came into prominence since they are guaranteed to have core stable solutions that can be found efficiently (Farrell and Scotchmer Q. J. Econ. 103(2), 279–297 1988). We improve upon existing results by proving that every instance of HGCRP has a solution that is Pareto optimal, core stable, and individually stable. The economic significance of this result is that efficiency is not to be totally sacrificed for the sake of stability in HGCRP. We establish that finding such a solution is NP-hard even if the sizes of the coalitions are bounded above by 3; however, it is polynomial time solvable if the sizes of the coalitions are bounded above by 2. We show that the gap between the total utility of a core stable solution and that of the socially-optimal solution (OPT) is bounded above by n, where n is the number of agents, and that this bound is tight. Our investigations reveal that computing OPT is inapproximable within better than (O(n^{1-epsilon })) for any fixed (epsilon > 0), and that this inapproximability lower bound is polynomially tight. However, OPT can be computed in polynomial time if the sizes of the coalitions are bounded above by 2.

A solver’s runtime and the quality of the solutions it generates are strongly influenced by its parameter settings. Finding good parameter configurations is a formidable challenge, even for fixed problem instance distributions. However, when the instance distribution can change over time, a once effective configuration may no longer provide adequate performance. Realtime algorithm configuration (RAC) offers assistance in finding high-quality configurations for such distributions by automatically adjusting the configurations it recommends based on instances seen so far. Existing RAC methods treat the solver as a black box, meaning the solver is given a configuration as input, and it outputs either a solution or runtime as an objective function for the configurator. However, analyzing intermediate output from the solver can enable configurators to avoid wasting time on poorly performing configurations. We propose a gray-box approach that utilizes intermediate output during evaluation and implement it within the RAC method Contextual Preselection with Plackett-Luce (CPPL blue). We apply cost-sensitive machine learning with pairwise comparisons to determine whether ongoing evaluations can be terminated to free resources. We compare our approach to a black-box equivalent on several experimental settings and show that our approach reduces the total solving time in several scenarios and improves solution quality in an additional scenario.