The psychology of robots

In recent years, neural networks have been proposed that portray many of the complexities of adaptive behavior. The networks describe how agents learn to predict future events by: 1) building models of the would, 2) inferring new predictions from past experiences, 3) combining elementary environmental stimuli into complex internal representations, 4) attending to stimuli associated with environmental novelty, and 5) attending to stimuli that are good predictors of other environmental events. When a predictive network is attached to a goal seeking system, the resulting architecture is able to describe spatial and maze navigation, as well as problem solving and planning. When the predictions of future events are based on the combination of environmental stimuli and the animal's own responses the networks provide the information necessary to choose between alternative behaviors. When the agent's own responses can be identified with the responses of other agents, the networks can describe learning by imitation. It is suggested that these principles might be applied to the design of adaptive, communicating autonomous robots     

A Comprehensive Survey of Multiagent Reinforcement Learning

Multiagent systems are rapidly finding applications in a variety of domains, including robotics, distributed control, telecommunications, and economics. The complexity of many tasks arising in these domains makes them difficult to solve with preprogrammed agent behaviors. The agents must, instead, discover a solution on their own, using learning. A significant part of the research on multiagent learning concerns reinforcement learning techniques. This paper provides a comprehensive survey of multiagent reinforcement learning (MARL). A central issue in the field is the formal statement of the multiagent learning goal. Different viewpoints on this issue have led to the proposal of many different goals, among which two focal points can be distinguished: stability of the agents' learning dynamics, and adaptation to the changing behavior of the other agents. The MARL algorithms described in the literature aim---either explicitly or implicitly---at one of these two goals or at a combination of both, in a fully cooperative, fully competitive, or more general setting. A representative selection of these algorithms is discussed in detail in this paper, together with the specific issues that arise in each category. Additionally, the benefits and challenges of MARL are described along with some of the problem domains where the MARL techniques have been applied. Finally, an outlook for the field is provided.     

Agents on the move [mobile software agents]

The author describes a new kind of software, based on artificial intelligence research, that can move itself from place to place to help people work more effectively. Known as agents, these artificial assistants are software components that live inside computer environments. Developed out of research in artificial intelligence (AI), agents were made in a variety of forms to perform all sorts of useful work-including obtaining airline departure dates and times, filtering e-mail for messages the user considers important, alerting users to significant stock price changes, and a host of other tasks. At first, agents were constrained to a single computer or at most to a single computing environment-a closed, homogenous network of, say, Unix platforms. Their behavior was limited and all the tasks they could do had to be pre-established. Today, agents are breaking the bonds that confine them to a single environment while learning new ways of accomplishing tasks on their own, based on their experience. The newcomers are called mobile agents, because they can move from one computer to another. As they emerge from the shadow of AI research, they are bringing together telecommunications, software, and distributed-system technologies to create new ways of getting things done     

Online pattern classification with multiple neural network systems: an experimental study

In this paper, an empirical study of the development and application of a committee of neural networks on online pattern classification tasks is presented. A multiple classifier framework is designed by adopting an Adaptive Resonance Theory-based (ART) autonomously learning neural network as the building block. A number of algorithms for combining outputs from multiple neural classifiers are considered, and two benchmark data sets have been used to evaluate the applicability of the proposed system. Different learning strategies coupling offline and online learning approaches, as well as different input pattern representation schemes, including the "ensemble" and "modular" methods, have been examined experimentally. Benefits and shortcomings of each approach are systematically analyzed and discussed. The results are comparable, and in some cases superior, with those from other classification algorithms. The experiments demonstrate the potentials of the proposed multiple neural network systems in offering an alternative to handle online pattern classification tasks in possibly nonstationary environments.     

基于自监督对比学习的深度神经网络对抗鲁棒性提升

基于深度神经网络的多源图像内容自动分析与目标识别方法近年来不断取得新的突破,并逐步在智能安防、医疗影像辅助诊断和自动驾驶等多个领域得到广泛部署。然而深度神经网络的对抗脆弱性给其在安全敏感领域的部署带来巨大安全隐患。对抗鲁棒性的有效提升方法是采用最大化网络损失的对抗样本重训练深度网络,但是现有的对抗训练过程生成对抗样本时需要类别标记信息,并且会大大降低无攻击数据集上的泛化性能。本文提出一种基于自监督对比学习的深度神经网络对抗鲁棒性提升方法,充分利用大量存在的无标记数据改善模型在对抗场景中的预测稳定性和泛化性。采用孪生网络架构,最大化训练样本与其无监督对抗样本间的多隐层表征相似性,增强模型的内在鲁棒性。本文所提方法可以用于预训练模型的鲁棒性提升,也可以与对抗训练相结合最大化模型的“预训练+微调”鲁棒性,在遥感图像场景分类数据集上的实验结果证明了所提方法的有效性和灵活性。      

Neural networks for frequency line tracking

This paper investigates the application of neural networks to frequency line tracking. Recently, hidden Markov models (HMM's) have been successfully applied to this problem, and here, we study a neural network architecture called Mnet, which is based on an underlying Markov model representation. A supervised learning algorithm is developed for Mnet, and a method of analytically deriving the connection weights for the Mnet is also mentioned. Two more conventional neural networks are also studied; a multilayer feedforward network and a multilayer network with feedback. The simulation results show that all three neural networks are comparable in performance to a hidden Markov model when applied to the frequency line tracking problem     

移动感知器网络中基于随机游走和协作关系的任务分发算法

关于移动感知器网络中感知任务的分发问题,目前学术界已经有了诸多相关研究.然而,这些研究很少涉及到多个智能体协作完成复杂感知任务问题.针对这种情况,首先,通过分析移动感知器网络的结构特征、智能体相互之间、以及智能体和感知任务之间的关系,本文提出了智能体之间协作关系强度和智能体对感知任务适应度两个概念,并讨论了二者对于移动感知器网络中感知任务动态分发的作用.其次,在上述概念的基础上,将二者融合为偏好因子,提出了基于随机游走和协作关系的任务分发算法（TDCR,Task Distribution With Cooperative Relationship）,通过该算法达到提高任务分发效率的目的.最后,将TDCR与Personal Rank算法（PR）、HITS算法对比分析,表明所提出的算法TDCR在任务分发效率和准确度等性能指标上有较好的提升.     

On the behavior-based architectures of autonomous agency

A number of autonomous robots with varying degrees of reactive functionality have been built, based on different architectures. We review the foundations, limitations, and achievements of a number of architectures of such autonomous agents from the three categories: (1) reactive; (2) deliberative; and (3) hybrid. Most of these architectures contain behaviors. The principle of avoiding an explicit representation of goals in the purely behavior-based robots has limited their achievements to simple tasks like box pushing, pipe inspection, and navigation. This paper makes two contributions: (1) reviewing autonomous agent architectures and (2) proposing a new class of architectures where behaviors are coupled and/or markers are introduced in environment, without a planner or sequencer and without an explicit representation of goals and investigating tradeoffs in these architectures. We develop a model of behaviors, environmental modification and goals and then show how the behavior-based robots can be made goal-directed. The tradeoffs in increasing their goal directedness are examined. Defining the notion of coupling that captures dependency within the internal structure of a behavior space, it is shown that more complex goals demand higher coupling or more behaviors or a modification to the environment. These novel tradeoffs show a new spectrum of architectures for integrating goals and the behavior-based reactive functionality.     

Applying Personal Construct Theory to Requirements Elicitation

One of the main approaches to Requirements Engineering is Goal-Oriented Requirement Engineering. This approach, based in Artificial Intelligence models, argues that goals are a natural and high level abstraction concept to elicit and represent requirements. Another advantage of goals is that they help non-functional requirements representation and reasoning. Although the goal oriented approach helps in representation and analysis, the problem of eliciting goals and their refinement is not trivial. In this article, we explore several applications to goal elicitation using a psychological theory: the Personal Construct Theory (PCT). This theory, stated by Kelly in 1955, can be used to elicit goals and their relationships. The choice of this theory is based in that it has a statistical base, therefore it is more precise and user independent than others. Finally, PCT can be amenable to automation by means of the Repertory Grid technique, widely studied in the Knowledge eliciting field.     

Distributed resource management in wireless sensor networks using reinforcement learning

In wireless sensor networks (WSNs), resource-constrained nodes are expected to operate in highly dynamic and often unattended environments. Hence, support for intelligent, autonomous, adaptive and distributed resource management is an essential ingredient of a middleware solution for developing scalable and dynamic WSN applications. In this article, we present a resource management framework based on a two-tier reinforcement learning scheme to enable autonomous self-learning and adaptive applications with inherent support for efficient resource management. Our design goal is to build a system with a bottom-up approach where each sensor node is responsible for its resource allocation and task selection. The first learning tier (micro-learning) allows individual sensor nodes to self-schedule their tasks by using only local information, thus enabling a timely adaptation. The second learning tier (macro-learning) governs the micro-learners by tuning their operating parameters so as to guide the system towards a global application-specific optimization goal (e.g., maximizing the network lifetime). The effectiveness of our framework is exemplified by means of a target tracking application built on top of it. Finally, the performance of our scheme is compared against other existing approaches by simulation. We show that our two-tier reinforcement learning scheme is significantly more efficient than traditional approaches to resource management while fulfilling the application requirements.     

面向复杂网络的异构网络表示学习综述

异构信息网络包含丰富的节点信息和链接信息,具有复杂异质性、高稀疏性、属性高维性等特性,这些特性给网络表示学习任务带来了巨大的挑战。异构网络表示学习通过在嵌入过程中将多样化的异质信息和结构信息进行有效融合,学习得到更有利于下游机器学习任务的低维特征向量。从异构网络表示学习方法的研究粒度出发,对近年的研究现状进行了比较全面的分析和讨论。首先探讨网络表示学习的产生动机,阐述了近年的异构网络表示学习的研究历程;然后对具有代表性的算法模型进行分类讨论,归纳其主要的研究内容和所使用的嵌入技巧。最后给出了未来工作中异构网络表示学习可能的研究方向和比较有价值的研究内容。     

动态网络表示学习研究进展

网络表示学习旨在将网络信息表示为低维稠密的实数向量,解决链接预测、异常检测、推荐系统等任务.近年来,网络表示学习研究取得重大进展,但研究多基于静态网络,而真实世界构成的网络是动态变化的,对动态网络分析的需求日益增加.本文总结了当前动态网络表示学习的方法与研究进展,首先提出网络表示学习的动机,阐述动态网络以及表示学习的发展历史与理论基础;接着,系统概述了大量动态网络嵌入方法,包括基于矩阵分解的动态图嵌入、基于随机游走的动态图嵌入、基于深度学习的动态图嵌入和基于重构概率的动态图嵌入,并分析与比较,给出动态网络表示学习的应用场景;最后,总结未来网络表示学习的研究方向.只有考虑网络的动态性,才能真实反映现实网络的演化,使网络表示学习更具价值.     

A Unified Artificial Neural Network Architecture for Active Power Filters

In this paper, an efficient and reliable neural active power filter (APF) to estimate and compensate for harmonic distortions from an AC line is proposed. The proposed filter is completely based on Adaline neural networks which are organized in different independent blocks. We introduce a neural method based on Adalines for the online extraction of the voltage components to recover a balanced and equilibrated voltage system, and three different methods for harmonic filtering. These three methods efficiently separate the fundamental harmonic from the distortion harmonics of the measured currents. According to either the Instantaneous Power Theory or to the Fourier series analysis of the currents, each of these methods are based on a specific decomposition. The original decomposition of the currents or of the powers then allows defining the architecture and the inputs of Adaline neural networks. Different learning schemes are then used to control the inverter to inject elaborated reference currents in the power system. Results obtained by simulation and their real-time validation in experiments are presented to compare the compensation methods. By their learning capabilities, artificial neural networks are able to take into account time-varying parameters, and thus appreciably improve the performance of traditional compensating methods. The effectiveness of the algorithms is demonstrated in their application to harmonics compensation in power systems     

Conditional distribution learning with neural networks and itsapplication to channel equalization

We present a conditional distribution learning formulation for real-time signal processing with neural networks based on an extension of maximum likelihood theory-partial likelihood (PL) estimation-which allows for (i) dependent observations and (ii) sequential processing. For a general neural network conditional distribution model, we establish a fundamental information-theoretic connection, the equivalence of maximum PL estimation, and accumulated relative entropy (ARE) minimization, and obtain large sample properties of PL for the general case of dependent observations. As an example, the binary case with the sigmoidal perceptron as the probability model is presented. It is shown that the single and multilayer perceptron (MLP) models satisfy conditions for the equivalence of the two cost functions: ARE and negative log partial likelihood. The practical issue of their gradient descent minimization is then studied within the well-formed cost functions framework. It is shown that these are well-formed cost functions for networks without hidden units; hence, their gradient descent minimization is guaranteed to converge to a solution if one exists on such networks. The formulation is applied to adaptive channel equalization, and simulation results are presented to show the ability of the least relative entropy equalizer to realize complex decision boundaries and to recover during training from convergence at the wrong extreme in cases where the mean square error-based MLP equalizer cannot     

基于子波分解的多通道神经网络纹理分割方法

描述了一种体现多通道滤波技术的神经网络纹理分割方法，决策神经网络（ＤＢＮ）可提高纹理分类的精度，同时纹理的子波变换降低了图像数据间的相关性，提高了网络的学习效率，实验表明本文提出孤方法分类误差较低，获得了令人满意的纹理分割效果。     

Intelligent tutoring interoperability for the new web

Interoperability of systems based on knowledge is a very important element for reducing their development cost and enabling an easy-to-perform service enrichment. Intelligent tutoring systems (ITSs) may be described as distant learning systems, which base their work on the simulation of the “real” teacher in the learning and teaching process. ITSs base their interoperability on the interchange of domain knowledge, knowledge about learning and teaching process and knowledge about students. This paper describes DiSNeT, a distance learning system we designed based on the intelligent tutoring paradigm, on knowledge presentation using distributed semantic networks and on using agents in the learning and teaching process. We also present a methodology for ensuring interoperability between DiSNeT and other ITSs.     

Boosted learning in dynamic Bayesian networks for multimodal speaker detection

Bayesian network models provide an attractive framework for multimodal sensor fusion. They combine an intuitive graphical representation with efficient algorithms for inference and learning. However, the unsupervised nature of standard parameter learning algorithms for Bayesian networks can lead to poor performance in classification tasks. We have developed a supervised learning framework for Bayesian networks, which is based on the Adaboost algorithm of Schapire and Freund. Our framework covers static and dynamic Bayesian networks with both discrete and continuous states. We have tested our framework in the context of a novel multimodal HCI application: a speech-based command and control interface for a Smart Kiosk. We provide experimental evidence for the utility of our boosted learning approach.     

The past, present, and future of neural networks for signalprocessing

The article provides a review of the fundamental of neural networks and reports recent progress. Topics covered include dynamic modeling, model-based neural networks, statistical learning, eigenstructure-based processing, active learning, and generalization capability. Current and potential applications of neural networks are also described in detail. Those applications include optical character recognition, speech recognition and synthesis, automobile and aircraft control, image analysis and neural vision, and several medical applications. Essentially, neural networks have become a very effective tool in signal processing, particularly in various recognition tasks     

Acoustic Event Detection in Multichannel Audio Using Gated Recurrent Neural Networks with High‐Resolution Spectral Features

Recently, deep recurrent neural networks have achieved great success in various machine learning tasks, and have also been applied for sound event detection. The detection of temporally overlapping sound events in realistic environments is much more challenging than in monophonic detection problems. In this paper, we present an approach to improve the accuracy of polyphonic sound event detection in multichannel audio based on gated recurrent neural networks in combination with auditory spectral features. In the proposed method, human hearing perception‐based spatial and spectral‐domain noise‐reduced harmonic features are extracted from multichannel audio and used as high‐resolution spectral inputs to train gated recurrent neural networks. This provides a fast and stable convergence rate compared to long short‐term memory recurrent neural networks. Our evaluation reveals that the proposed method outperforms the conventional approaches.     

基于高效用神经网络的文本分类方法

现有的基于深度学习的文本分类方法没有考虑文本特征的重要性和特征之间的关联关系,影响了分类的准确率.针对此问题,本文提出一种基于高效用神经网络（High Utility Neural Networks,HUNN）的文本分类模型,可以有效地表示文本特征的重要性及其关联关系.利用高效用项集挖掘（Mining High Utility Itemsets,MHUI）算法获取数据集中各个特征的重要性以及共现频率.其中,共现频率在一定程度上反映了特征之间的关联关系.将MHUI作为HUNN的挖掘层,用于挖掘每个类别数据中重要性和关联性强的文本特征.然后将这些特征作为神经网络的输入,再经过卷积层进一步提炼类别表达能力更强的高层次文本特征,从而提高模型分类的准确率.通过在6个公开的基准数据集上进行实验分析,提出的算法优于卷积神经网络（Convolutional Neural Networks,CNN）,循环神经网络（Recurrent Neural Networks,RNN）,循环卷积神经网络（Recurrent Convolutional Neural Networks,RCNN）,快速文本分类（Fast Text Classifier,FAST）,分层注意力网络（Hierarchical Attention Networks,HAN）等5个基准算法.