PaperlayerView exact revision on arXiv

Source-linked AI summary

Cognitive Internet of Things: A New Paradigm beyond Connection

Qihui Wu, Guoru Ding, Yuhua Xu, Shuo Feng, Zhiyong Du, Jinlong Wang, Keping Long

arXiv:1403.2498v1cs.AI

TL;DR

The paper addresses the limitation of IoT systems that primarily connect objects without enabling them to learn, think, and understand physical and social worlds. It defines CIoT through human cognition, develops an operational framework around five cognitive tasks, and surveys enabling techniques and evaluation metrics. The paper concludes that CIoT can bridge physical and social worlds and support smarter allocation, network operation, and service provisioning, while identifying unresolved convergence and theory-to-practice challenges.

  • Problem

    Existing IoT mainly connects objects and shares observations, while general objects still lack capabilities to learn, think, and understand physical and social worlds independently.

  • Method

    The paper defines CIoT by integrating human cognitive processes into IoT system design and develops a framework covering five cognitive tasks, enabling techniques, and performance metrics.

  • Results

    The paper presents CIoT as a paradigm in which general objects can learn, think, and understand physical and social worlds, supported by an operational cognitive framework.

  • Takeaways & Limitations

    CIoT has the capability to bridge physical and social worlds and enhance smart resource allocation, automatic network operation, and intelligent service provisioning.

  • Takeaways & Limitations

    The paper identifies knowledge-assisted learning for faster convergence and stronger theory-to-practice links as open research challenges.

Abstract

from arXiv · show

Current research on Internet of Things (IoT) mainly focuses on how to enable general objects to see, hear, and smell the physical world for themselves, and make them connected to share the observations. In this paper, we argue that only connected is not enough, beyond that, general objects should have the capability to learn, think, and understand both physical and social worlds by themselves. This practical need impels us to develop a new paradigm, named Cognitive Internet of Things (CIoT), to empower the current IoT with a `brain' for high-level intelligence. Specifically, we first present a comprehensive definition for CIoT, primarily inspired by the effectiveness of human cognition. Then, we propose an operational framework of CIoT, which mainly characterizes the interactions among five fundamental cognitive tasks: perception-action cycle, massive data analytics, semantic derivation and knowledge discovery, intelligent decision-making, and on-demand service provisioning. Furthermore, we provide a systematic tutorial on key enabling techniques involved in the cognitive tasks. In addition, we also discuss the design of proper performance metrics on evaluating the enabling techniques. Last but not least, we present the research challenges and open issues ahead. Building on the present work and potentially fruitful future studies, CIoT has the capability to bridge the physical world (with objects, resources, etc.) and the social world (with human demand, social behavior, etc.), and enhance smart resource allocation, automatic network operation, and intelligent service provisioning.

I. INTRODUCTION

The paper introduces Cognitive Internet of Things (CIoT) to extend IoT beyond connectivity, enabling objects to learn, think, and understand physical and social worlds with minimal human intervention. It defines CIoT, motivates its physical-cyber-social integration, and distinguishes it from related IoT and cognitive-radio paradigms.

  • A. Background and Motivation: The paper presents application scenarios in which connected objects sense users or traffic conditions, share observations, and support more intelligent services.Examples include coordinated smart-home devices and traffic information gathered from cameras, vehicles, drivers, and passengers.
  • A. Background and Motivation: The paper motivates CIoT because existing IoT applications remain highly dependent on humans for cognition processing despite increasing interconnectivity.The authors characterize connectivity alone as insufficient for realizing expected intelligent and convenient applications.
  • A. Background and Motivation: CIoT is proposed as a new paradigm that gives interconnected physical or virtual objects cognitive capabilities and lets them behave as agents with minimal human intervention.Its operational definition combines context-aware perception-action cycles, learning from physical and social environments, semantic or knowledge storage, and adaptive decision-making.
  • A. Background and Motivation: CIoT aims to bridge the physical and social worlds with objects and resources on one side and human demand and social behavior on the other.This integration forms an intelligent physical-cyber-social system.
  • B. Historical Notes: CIoT differs from cognitive radio networks because it targets massive heterogeneous general objects and varied application objectives rather than radio-spectrum utilization and wireless throughput.Both paradigms draw on cognitive science, but their systems and objectives are substantially different.

C. Purpose of this Paper

The paper introduces Cognitive Internet of Things (CIoT) to extend IoT objects with cognition that learns, thinks, and understands physical and social worlds. It defines a framework connecting physical environments and social networks through five cognitive tasks and discusses enabling techniques and applications.

  • Purpose of this Paper: CIoT integrates the operational process of human cognition into IoT design so general objects can learn, think, and understand physical and social worlds.
  • Applications: The framework and enabling techniques can support applications such as intelligent transportation systems and software or Apps from CIoT companies.
  • Framework of CIoT: The framework bridges physical things, objects, and resources with human demand and social behavior as an intelligent physical-cyber-social system.
  • Framework of CIoT: The sensing-control layer interfaces with the physical environment, while the data-semantic-knowledge and decision-making layers analyze information and select actions.
  • Framework of CIoT: The service evaluation layer provisions on-demand services to social networks and feeds performance results back into the cognition process.
  • Fundamental Cognitive Tasks: Its five cognitive tasks are perception-action, massive data analytics, semantic derivation and knowledge discovery, intelligent decision-making, and on-demand service provisioning.

III. MASSIVE DATA ANALYTICS IN COGNITIVE INTERNET OF THINGS

Massive data analytics is needed because future CIoT systems generate noisy, corrupted, heterogeneous, high-dimensional, and nonlinear data. The paper organizes enabling algorithms into four classes and recommends copula-based processing for heterogeneous data.

  • Massive Data Analytics: CIoT data from cameras, vehicles, drivers, and passengers can be noisy, corrupted, heterogeneous, high-dimensional, and nonlinearly separable.
  • Massive Data Analytics: Effective massive-data algorithms are urgently needed to analyze, interpret, understand, and exploit the value of collected data.
  • Massive Data Analytics: The tutorial groups massive-data analytics into heterogeneous, nonlinear, high-dimensional, and distributed or parallel data processing.
  • Heterogeneous Data Processing: Heterogeneous sensors produce modalities such as text, video, and voice, creating processing challenges while offering possibilities for system improvement.
  • Heterogeneous Data Processing: The resulting joint probability density can support multi-sensor fusion, distribution estimation, and distributed detection.
  • Heterogeneous Data Processing: Copula theory couples multivariate joint distributions to marginal distributions, avoiding potentially suboptimal simple product or multivariate Gaussian assumptions.

B. Nonlinear Data Processing

The paper motivates nonlinear data processing because optimal fusion rules are generally nonlinear and linear methods can be oversimplified. Kernel-based learning supplies nonlinear variants of established linear statistical techniques.

  • Nonlinear Data Processing: Optimal fusion rules derived from multivariate joint distributions are generally mathematically intractable because they involve nonlinear operations.
  • Nonlinear Data Processing: Linear data-processing methods dominate development for simplicity, but they can deviate from optimality through oversimplification.
  • Nonlinear Data Processing: Kernel-based learning constructs powerful nonlinear variants of widely used statistical linear techniques.
  • Nonlinear Data Processing: KBL maps input data into a higher-dimensional feature space and replaces explicit inner products with an appropriate positive semi-definite symmetric kernel.
  • Nonlinear Data Processing: Projective kernels depend on inner products, whereas radial kernels depend on distances, including polynomial and Gaussian examples.

C. High-Dimensional Data Processing

High-dimensional CIoT data often has intrinsic low-rank structure, motivating matrix recovery and completion methods. The paper progresses from recovery with noise to sparse anomalies and incomplete observations.

  • High-Dimensional Data Processing: Images and videos in CIoT are high-dimensional, while kernel mappings can further increase the effective dimensionality.
  • High-Dimensional Data Processing: Traffic and video-surveillance matrices can exhibit intrinsic low rank because of shared temporal patterns and periodic behavior.
  • High-Dimensional Data Processing: Matrix completion recovers missing entries, whereas matrix recovery reconstructs matrices containing corrupted entries.
  • Low-Rank Matrix Recovery: For noisy sensing data, PCA-style formulations recover an unknown low-rank matrix from a matrix modeled as low-rank structure plus perturbation.
  • Low-Rank Matrix Recovery: Sparse anomalies can be modeled separately from low-rank data and recovered through tractable convex optimization using a rank-sparsity parameter and l1-norm.
  • Joint Matrix Completion and Recovery: Transmission loss and limited crowdsourcing incentives make sensing matrices noisy, corrupted, and incomplete, requiring projection onto acquired entries for joint completion and recovery.
  • High-Dimensional Data Processing: Augmented Lagrange multipliers and accelerated proximal gradient algorithms are identified as existing approaches for efficiently solving the recovery and completion problems.

D. Parallel and Distributed Data Processing

Centralized processing can be inefficient or impossible for decentralized CIoT networks, motivating parallel and distributed approaches. ADMM provides a framework that coordinates local computations with global or neighbor-based aggregation.

  • Decentralized CIoT applications challenge centralized processing because of single-node failure, limited scalability, and large exchange overhead.
  • ADMM reformulates a common global objective using local variables and a shared global variable, producing a global consensus problem.
  • The resulting ADMM algorithm separates independent local updates from a global aggregation step.
  • For multi-hop smart-object networks, aggregation can be replaced by local fusion of one-hop neighbors’ data.
  • ADMM can coordinate existing smaller-scale model-fitting algorithms to solve larger distributed problems.
  • Each ADMM iteration can be represented in MapReduce, with Maps performing local computations and Reduce performing global aggregation.

IV. SEMANTIC DERIVATION AND KNOWLEDGE DISCOVERY IN COGNITIVE INTERNET OF THINGS

CIoT must transform organized data from physical, cyber, and social worlds into semantic meaning and reusable knowledge. Semantic derivation addresses heterogeneous contexts, while knowledge discovery extracts patterns and rules for intelligent behavior.

  • Analyzed CIoT data organize observations from physical, cyber, and social worlds but remain too complex and inefficient for direct decision-making.
  • Semantic derivation automatically obtains meaning from analyzed data so CIoT objects can understand and become aware.
  • Knowledge discovery identifies valuable patterns or rules from analyzed data and derived semantic.
  • In CIoT, context represents analyzed data organized to describe an entity’s situation, including location, identity, time, or activity.
  • Massive heterogeneous context sources make identical situations difficult to understand because they may express the same semantic in different forms.
  • Ontology-based semantic technologies help combine different contexts and obtain their shared meaning.

2) Ontology in CIoT:

Ontology provides CIoT with a formal representation for shared concepts, relations, and context across heterogeneous sources. Semantic standardization extends this interoperability, while knowledge discovery converts analyzed data into more stable, useful knowledge.

  • Ontology in CIoT: In CIoT, an ontology formally represents relations among terms through a shared conceptualization of a phenomenon.
  • Ontology in CIoT: Ontology expresses relationships and helps identify identical semantics originating from different contexts.
  • Semantic standardization: Semantic standardization increases interoperability and extendibility across massive heterogeneous data and context sources.
  • Semantic standardization: Standards enable connected objects to continuously share and exchange semantic information across diverse systems.
  • CIoT’s high-level intelligence must be addressed throughout design, development, implementation, and evaluation rather than in one isolated component.
  • Knowledge discovery from analyzed data is identified as a primary route to achieving intelligence and applying it subsequently.
  • Knowledge is described as valid, certain, potentially useful, consolidated, contextualized, and more stable over time than data, context, or semantic.
  • Artificial intelligence, machine learning, pattern recognition, and database techniques can discover knowledge from analyzed data.

1) Association Analysis:

CIoT decision-making selects actions using semantic information, knowledge, service demand, and sometimes information about other decision-makers. Because selections interact, the paper emphasizes multi-agent learning and game-theoretic models under varied information conditions.

  • Decision-making: Decision-making in CIoT is treated as selecting among alternatives such as traffic paths, wireless channels, and services.
  • Decision-making: The paper focuses on multi-agent learning because CIoT is expected to contain many interactive human or machine decision-makers.
  • Decision-making: A decision-maker combines environmental information, available information about others, and service demand to produce a decision.
  • Game-theoretic learning: Game models capture interactive selections, while non-cooperative games represent distributed autonomous decision-making through individual utility maximization.
  • Game-theoretic learning: Game-theoretic learning must design utility functions and achieve stable solutions under static or dynamic systems and complete or incomplete information.

B. Intelligent Decision-Making in Large-Scale CIoT

Large-scale CIoT decision-making involves many spatially distributed agents with local interactions and uncertain, dynamic, or incomplete information. The paper frames these decisions as games and uses learning procedures to update selections toward stable outcomes.

  • Spatial interactions: CIoT decisions involve many spatially distributed agents whose selections directly affect nearby decision-makers.Examples include nearby vehicles and neighboring sensors competing for channels, energy, or time.
  • Spatial interactions: Spatial games model each player’s interactive neighbors, selection set, and utility function, making them more suitable than global games for large-scale CIoT.The local interaction structure is represented by G_l = {N, J_n, A_n, u_n}.
  • Information constraints: CIoT decision-making must account for uncertain vehicle arrivals, dynamic congestion, and incomplete sensor information.These constraints are illustrated through smart traffic systems.
  • Information constraints: The main optimization objectives are maximizing expected payoff or minimizing outage probability under constrained information.The corresponding forms are max E[r_n(t)] and min Pr{r_n(t) > η_n}.
  • Learning procedure: Learning updates each player’s next selection from its current selection and random payoff, repeating until a stopping criterion is reached.The update rule is a_n(t + 1) = F[a_n(t), r_n(t)].
  • Performance evaluation: Evaluating CIoT service performance is challenging because many considerations and factors must be incorporated into the metric structure.The considered structure includes profit-related layers such as data, information, and user quality.

1) Data layer-QoD:

The proposed evaluation framework measures CIoT performance through data, information, user experience, and resource-efficiency dimensions. It links QoD, QoI, and QoE to successive layers while also accounting for device, computational, energy, and storage costs.

  • Data layer-QoD: The data-layer metric evaluates sensed-data quality, acquisition, and distribution during the perception and sensing stage.The paper treats data quality as fundamental to the CIoT cycle.
  • Data layer-QoD: IoT data may be noisy, device-dependent, maliciously corrupted, or incomplete because of sensing limitations.These factors can prevent acquired data from meeting system requirements.
  • Information layer-QoI: QoI evaluates whether information supplied to decision-making meets the decision-maker’s needs in a specific place, location, and social setting.The framework treats information as the input to intelligent decision-making.
  • Information layer-QoI: QoI combines quantity, precision, recall, accuracy, detail, timeliness, and validity, with normalized values from 0 to 1.A value of 1 represents the corresponding best case.
  • User layer-QoE: QoE measures user profit as the subjective acceptability of IoT applications or services, extending beyond communication quality to upper-layer factors.The framework evaluates QoE across four service-provisioning levels and allows mappings to reflect heterogeneous user demands.
  • Cost dimension: The cost dimension measures resource efficiency across device utilization, computation, energy, and storage.These costs arise from constrained hardware, computational load, energy use across cognitive tasks, and growing storage demand.

VII. RESEARCH CHALLENGES AND OPEN ISSUES

The paper identifies major open issues separating CIoT’s proposed vision from current research and development. Challenges span heterogeneous data processing, multi-agent decision-making, knowledge-assisted learning, system-level QoE, application-specific constraints, and theory-to-practice translation.

  • Overall scope: Current CIoT research and development remain far from the paper’s envisioned future, leaving several major challenges and open issues.The authors present the listed challenges as non-exhaustive.
  • Data analytics: Processing massive sensing data remains challenging when data are heterogeneous, high-dimensional, and nonlinearly separable.These mixed characteristics occur in practical CIoT applications.
  • Decision-making: Large-scale CIoT requires game models and multi-agent learning algorithms that account for application differences and practical decision-making constraints.The constraints include local interaction and uncertain, dynamic, and incomplete information.
  • Decision-making: Existing history-based multi-agent learning may converge slowly because players must explore possible selections, motivating knowledge-assisted learning technologies.The proposed direction seeks faster convergence and better performance.
  • Semantics and knowledge: Effective semantic technologies and knowledge-discovery techniques tailored to CIoT applications remain a fundamental research task.The challenge concerns suitability for practical CIoT applications.
  • Performance evaluation: QoE research is mostly limited to single-user cases, with insufficient system-level evaluation for large-scale CIoT systems with massive users.The limitation concerns the current scope of QoE studies.
  • Theory and practice: Generic CIoT approaches require additional practical constraints before they can be applied to specific situations, and theory must be connected to practice.The authors ask how theoretical studies can be applied, including by companies implementing smart cities.
Loading 1403.2498v1…