Source-linked AI summary
Enabling Cognitive Smart Cities Using Big Data and Machine Learning: Approaches and Challenges
Mehdi Mohammadi, Ala Al-Fuqaha
TL;DR
Smart cities generate enormous, rapidly changing datasets, much of which remains unused because labels and established data-use mechanisms are scarce. The paper proposes a hierarchical framework using semi-supervised deep reinforcement learning, and reports improved rewards and localization accuracy over supervised learning.
Problem
Smart city data is underutilized because enormous data volumes include much unlabeled information and lack established mechanisms for extracting knowledge.
Method
The paper combines semi-supervised learning and deep reinforcement learning within a three-level intelligence framework spanning IoT, fog, and cloud environments.
Results
Semi-supervised DRL achieved 60% to 100% more total rewards and 6% to 23% better localization accuracy than supervised DRL in the reported comparison.
Takeaways & Limitations
The framework is presented as an enabler for cognitive smart-city services that adapt as environmental conditions change and perform autonomic actions.
Abstract
from arXiv · showhide
The development of smart cities and their fast-paced deployment is resulting in the generation of large quantities of data at unprecedented rates. Unfortunately, most of the generated data is wasted without extracting potentially useful information and knowledge because of the lack of established mechanisms and standards that benefit from the availability of such data. Moreover, the high dynamical nature of smart cities calls for new generation of machine learning approaches that are flexible and adaptable to cope with the dynamicity of data to perform analytics and learn from real-time data. In this article, we shed the light on the challenge of under utilizing the big data generated by smart cities from a machine learning perspective. Especially, we present the phenomenon of wasting unlabeled data. We argue that semi-supervision is a must for smart city to address this challenge. We also propose a three-level learning framework for smart cities that matches the hierarchical nature of big data generated by smart cities with a goal of providing different levels of knowledge abstractions. The proposed framework is scalable to meet the needs of smart city services. Fundamentally, the framework benefits from semi-supervised deep reinforcement learning where a small amount of data that has users' feedback serves as labeled data while a larger amount is without such users' feedback serves as unlabeled data. This paper also explores how deep reinforcement learning and its shift toward semi-supervision can handle the cognitive side of smart city services and improve their performance by providing several use cases spanning the different domains of smart cities. We also highlight several challenges as well as promising future research directions for incorporating machine learning and high-level intelligence into smart city services.
I. INTRODUCTION
Smart cities generate vast, dynamic, noisy data, but much remains unused because labeling and analytics mechanisms are insufficient. The paper proposes combining deep neural networks, reinforcement learning, and semi-supervised learning to extract abstractions and support adaptive autonomous services.
- Data and learning challenges: Smart city services generate data as producers, consumers, or both through deployed sensors, actuators, smart objects, and connected vehicles.Parking brokers publish availability data, while cars consume and may also produce data for other components.
- Data and learning challenges: Only a small fraction of massive smart city data is utilized, largely because insufficient labeled data limits machine learning exploitation of labeled and unlabeled data.The paper frames this underutilization as a data-recycling challenge.
- Data and learning challenges: Sampling can neglect important patterns, while deep neural networks can process datasets with millions of parameters to extract insightful analytics.The paper contrasts sampling efficiency with the broader data coverage enabled by DNNs.
- Data and learning challenges: By 2012, about 0.5% of 2.8 ZB of stored data had been analyzed and 3% had been labeled, highlighting potentially wasted information in 99.5% of generated data.The passage attributes these estimates to an IDC study and describes the evidence as anecdotal.
- Data and learning challenges: Smart city learning must operate continuously and dynamically because applications evolve, data arrives rapidly, and users cannot provide feedback for every training instance.Generated data may also be noisy or uncertain.
- Proposed direction: The proposed integration of DNNs, reinforcement learning, and semi-supervised learning delivers agents that learn from smart city data, adapt to changing environments, and act autonomously.DNNs provide high-level abstractions, while reinforcement learning selects actions using rewards and semi-supervision addresses partially labeled data.
II. RELATED WORK
Related work spans cognitive services, IoT intelligence, smart-home learning, and fog-based analytics. The paper positions its approach across fog and cloud levels while using untapped unlabeled data.
- Cognitive and IoT frameworks: Cognitive systems combine IoT technologies, generated big data, and artificial intelligence, with dynamic learning based on feedback from previous rounds.The passage also describes cognitive systems as capable of reasoning, interaction, and experience-based learning.
- Cognitive and IoT frameworks: Prior cognitive IoT frameworks address object interaction, massive analytics, semantic knowledge discovery, intelligent decisions, and on-demand service provisioning.Examples include layered virtual-object services and five interacting cognitive tasks.
- Cognitive and IoT frameworks: A smart-home cognitive framework combines Bayesian modeling, Bayesian filtering, and reinforcement learning to estimate system state and select actions from received rewards.The Bayesian model observes the environment through a perceptual component.
- Fog and cloud analytics: Fog-based intelligence has been proposed as an alternative to centralized cloud analytics, reducing communication bandwidth by avoiding transmission of all raw data.One cited model uses hierarchical fog computing and hidden Markov models for pipeline monitoring.
- Fog and cloud analytics: This study targets analytics on both fog and cloud levels and uses the larger untapped body of unlabeled data to improve analytic accuracy across time-sensitive and non-time-sensitive applications.Table I situates the study relative to prior work by showing covered intelligence and analytics levels.
III. INTELLIGENCE FOR SMART CITIES
The framework organizes smart-city intelligence across IoT infrastructure, fog computing, and cloud computing. Each level uses models suited to its resources and supports progressively broader data abstractions and actions.
- Framework hierarchy: The three intelligence levels are smart city and IoT infrastructure, fog computing, and cloud computing, with intelligent agents deployed according to analytics characteristics such as time sensitivity.The hierarchy matches machine learning placement to smart-city infrastructure.
- Framework hierarchy: As data moves through the infrastructure, it supports deeper abstractions and knowledge representations, from short-term local management to long-term city-wide resource management.Fog analytics support local predefined contexts, whereas cloud analytics cover larger geographical regions and varied contexts.
- IoT infrastructure: Resource-constrained IoT devices favor shallow unsupervised or semi-supervised models, while compressed or pruned neural networks can enable deeper models and deep reinforcement learning near data sources.The passage identifies model compression and pruning as approaches for devices, wearables, and smartphones.
- Fog computing: At the fog level, aggregated raw data moves toward the cloud, while compressed deep learning, deep reinforcement learning, and semi-supervised methods operate under fewer resource constraints than IoT devices.Lightweight intelligence is also needed at gateways and proxies.
- Cloud computing: At the cloud level, large-scale machine learning and data-mining frameworks can combine with semantic learning and ontologies to extract high-level insights and patterns.Deep learning is suited to this level because it provides deeper data abstractions.
IV. EMERGING APPROACHES
Deep reinforcement learning combines neural-network representations with reward-based action selection, while the proposed semi-supervised variant incorporates unlabeled data for adaptive smart-city services.
- Deep reinforcement learning uses a neural network to approximate the optimal action-value function for selecting actions that maximize future rewards.The agent observes environmental parameters, acts, and receives reward feedback.
- The proposed model adapts generative deep neural networks such as variational autoencoders to produce action probabilities as its semi-supervised component.
- In a smart-campus localization task, the agent learned navigation actions from Bluetooth RSSI fingerprints with 15% of data points labeled by location.The reward function was defined as the reciprocal of distance to the target.
- 60% to 100% more rewards and 6% to 23% better localization accuracy were achieved by semi-supervised DRL than supervised DRL across different epoch counts.The model reached the target more closely and obtained higher rewards more quickly.
V. SMART CITY USE CASES
The paper presents smart-city use cases illustrating semi-supervised learning for improving services to city residents, including performance comparisons between semi-supervised and supervised DRL.
- Smart-city services span resource management and control alongside transportation, healthcare, convenience, agriculture, and government applications.
- Figure 4 compares semi-supervised and supervised DRL using average total rewards and average distance to the target across epoch counts.
- The use cases illustrate semi-supervised learning as an approach for providing better services to city residents.
A. Water
The water use case combines sensor, forecast, meter, clustering, and DRL-based monitoring to support drought response, consumption analysis, and leak detection.
- Smart-city analytics can combine temperature and humidity sensors, weather forecasts, water-usage predictions, and resource data to help secure water during droughts.Crowd-sensing can also monitor creek water levels and quality.
- Smart water meters enable fine-grained household and city-level monitoring, while unsupervised clustering can detect abnormal consumption and leaks.
- A DRL system can learn a policy to stop tap water flow and notify a household when observed consumption deviates from its learned pattern.The example pattern is 20 ± 2 liters between 5 and 6 p.m. on weekdays.
- 12% reduction in water consumption in one year followed smart-meter installation that enabled early leak detection in Kalgoorlie-Boulder, Australia.
B. Energy
The energy use case applies semi-supervised DRL to smart-home and non-intrusive load monitoring, while highlighting equipment trade-offs and broader smart-grid benefits.
- Residential electricity use and peak-load management motivate smart meters that monitor household energy profiles and provide feedback.The residential sector consumes around one third of electricity usage in the European Union, and demand is predicted to double in the next decade.
- A semi-supervised DRL agent can use appliance profiles, temperature, light, and motion data to learn policies for turning devices off.Appliance Load Monitoring provides fine-grained device measurements but adds equipment cost and complexity.
- Non-Intrusive Load Monitoring extracts individual appliance usage from one aggregated household measurement, reducing the need for device-level instrumentation.The approach is trained using individual-appliance consumption data, events, and timestamps.
- The semi-supervised DRL agent performs better than the supervised DRL agent in the presence of unlabeled NILM data.
- $1 million was saved after a Chattanooga smart-grid incident through faster repairs following a severe storm outage.
C. Agriculture
Smart agriculture uses sensor-driven monitoring and automated corrective actions to improve crop production, while image-based systems can identify plant diseases and crop anomalies.
- Monitoring soil moisture and minerals can trigger corrective actuator actions such as adding water or minerals.
- Decision-making processes connected to agricultural monitoring can increase crop productivity.
- Image-based classification systems can identify diseased plants by analyzing crop or leaf images.
- Farmers can use smart devices to detect anomalous fruits and crops through visual disease-recognition systems.
VI. CHALLENGES AND FUTURE DIRECTIONS
Smart city machine-learning systems face challenges involving real-time analytics, security and privacy, device constraints, dataset availability, and contextual reasoning.
- Integrating big and fast data analytics: Time-sensitive smart city applications require frameworks that combine big-data analytics with real-time or near-real-time stream processing.
- Preserving security and privacy: False Data Injection attacks can compromise the validity and trustworthiness of data-driven machine-learning systems.
- Preserving security and privacy: Smart city systems must preserve privacy because much of their data originates from individuals who may not want it publicly available.
- On-device intelligence: Resource-constrained devices require lightweight machine-learning algorithms for hard real-time intelligence.
- Big dataset shortage: Real-world datasets are not readily available for many smart city application domains, requiring confirmation with simulated big data.
- Context-awareness: Contextual information must be integrated with raw data to support faster, more accurate reasoning and actuation.
B. Future Directions
The paper identifies decentralized analytics, adaptive transfer, semantic integration, virtual objects, and natural interaction as future directions for smart city intelligence.
- Decentralized analytics: Sending raw IoT data to the cloud is not effective or scalable for smart city deployments, motivating analytics closer to fog and device layers.
- Adaptive learning: Changing feature distributions can require rebuilding trained models, while cross-platform transfer may avoid collecting new localization data.
- Semantic integration: Semantic technologies are needed to support smart city systems that interact with citizens and use social media data.
- Intelligent virtual objects: Virtual representations of physical objects can learn, decide, and act autonomously when combined with deep reinforcement learning.
- Natural interaction: Automatic speech recognition and natural language understanding provide convenient interaction for small mobile and wearable devices.
VII. CONCLUSION
The paper addresses the mismatch between static traditional machine-learning assumptions and evolving smart city environments by proposing a hierarchical semi-supervised deep reinforcement learning framework and surveying related challenges and directions.
- Traditional machine-learning techniques often assume fixed training models and static contexts, unlike smart city environments whose data evolve over time.
- The article reviews data recycling, efficient sampling, scalable models, and state-of-the-art approaches for smart city big-data analytics.
- The proposed hierarchical framework aligns machine-learning techniques with the hierarchy of big data in smart cities.
- The paper proposes semi-supervised deep reinforcement learning to address the identified challenges across smart city application domains.
- The article highlights challenges and research directions for incorporating machine learning and high-level intelligence into new smart city services.