idnits 2.17.1 draft-choi-icnrg-aiot-03.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- == The page length should not exceed 58 lines per page, but there was 1 longer page, the longest (page 1) being 60 lines Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- ** The document seems to lack a Security Considerations section. ** The document seems to lack an Authors' Addresses Section. ** There are 511 instances of too long lines in the document, the longest one being 5 characters in excess of 72. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year == Line 778 has weird spacing: '... of unsup...' -- The document date (May 19, 2020) is 1409 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- == Missing Reference: 'Kriegal' is mentioned on line 688, but not defined == Missing Reference: 'Hochreiter' is mentioned on line 755, but not defined == Missing Reference: 'Krizhevsky' is mentioned on line 774, but not defined == Missing Reference: 'Goodfellow' is mentioned on line 793, but not defined == Missing Reference: 'Venkataramani' is mentioned on line 833, but not defined == Unused Reference: 'Kriegel' is defined on line 997, but no explicit reference was found in the text Summary: 3 errors (**), 0 flaws (~~), 9 warnings (==), 1 comment (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 ICN Research Group J.K. Choi 3 Internet Draft J.S. Han 4 Intended status: Informational G.H. Lee 5 Expires: November 19, 2020 N.K. Kim 6 KAIST 7 May 19, 2020 9 Requirements and Challenges for User-level Service Managements of 10 IoT Network by utilizing Artificial Intelligence 11 draft-choi-icnrg-aiot-03 13 Abstract 15 This document describes the requirements and challenges to employ 16 artificial intelligence (AI) into the constraint Internet of Things 17 (IoT) service environment for embedding intelligence and increasing 18 efficiency. 20 The IoT service environment includes heterogeneous and multiple IoT 21 devices and systems that work together in a cooperative and 22 intelligent way to manage homes, buildings, and complex autonomous 23 systems. Therefore, it is becoming very essential to integrate IoT 24 and AI technologies to increase the synergy between them. However, 25 there are several limitations to achieve AI enabled IoT as the 26 availability of IoT devices is not always high, and IoT networks 27 cannot guarantee a certain level of performance in real-time 28 applications due to resource constraints. 30 This document intends to present a right direction to empower AI in 31 IoT for learning and analyzing the usage behaviors of IoT 32 devices/systems and human behaviors based on previous records and 33 experiences. With AI enabled IoT, the IoT service environment can be 34 intelligently managed in order to compensate for the unexpected 35 performance degradation often caused by abnormal situations. 37 Status of this Memo 39 This Internet-Draft is submitted in full conformance with the 40 provisions of BCP 78 and BCP 79. 42 Internet-Drafts are working documents of the Internet Engineering 43 Task Force (IETF), its areas, and its working groups. Note that 44 other groups may also distribute working documents as Internet-Drafts. 46 Internet-Drafts are draft documents valid for a maximum of six months 47 and may be updated, replaced, or obsoleted by other documents at any 48 time. It is inappropriate to use Internet-Drafts as reference 49 material or to cite them other than as "work in progress." 51 The list of current Internet-Drafts can be accessed at 52 http://www.ietf.org/ietf/1id-abstracts.txt 54 The list of Internet-Draft Shadow Directories can be accessed at 55 http://www.ietf.org/shadow.html 57 This Internet-Draft will expire on November 19, 2020. 59 Copyright Notice 61 Copyright (c) 2020 IETF Trust and the persons identified as the 62 document authors. All rights reserved. 64 This document is subject to BCP 78 and the IETF Trust's Legal 65 Provisions Relating to IETF Documents 66 (http://trustee.ietf.org/license-info) in effect on the date of 67 publication of this document. Please review these documents carefully, 68 as they describe your rights and restrictions with respect to this 69 document. Code Components extracted from this document must include 70 Simplified BSD License text as described in Section 4.e of the Trust 71 Legal Provisions and are provided without warranty as described in 72 the Simplified BSD License. 74 Table of Contents 76 1. Introduction ................................................ 3 77 2. Challenging Issues of IoT network 78 ............................ 6 79 2.1. Untrusted and incorrect IoT devices ..................... 6 80 2.2. Traffic burstiness of IoT network ....................... 6 81 2.3. Management overheads of heterogeneous IoT sensors 82 ........ 7 83 3. Overview of AI/ML-based IoT services ......................... 9 84 4. Requirements for AI/ML-based IoT services ................... 11 85 4.1. Requirements for AI/ML-based IoT data collection and delivery 87 ........................................................... 11 88 4.2. Requirements for intelligent and context-aware IoT services12 89 4.3. Requirements for applying AI/ML to IoT data ............ 14 90 4.3.1. Training AI/ML algorithm 91 .......................... 14 92 4.3.2. AI/ML inference in IoT application ................ 15 93 5. State of arts of the artificial intelligence/machine learning 94 technologies for IoT services 95 .................................. 15 97 5.1. Machine learning and artificial intelligence technologies 98 review ..................................................... 15 99 5.1.1. Supervised learning for IoT ....................... 16 100 5.1.2. Unsupervised learning for IoT ..................... 17 101 5.1.3. Reinforcement learning for IoT .................... 18 102 5.1.4. Neural Network based algorithms for IoT ........... 19 103 5.2. Technologies for lightweight and real-time intelligence 104 . 20 105 6. Use cases of AI/ML into IoT service ......................... 22 106 6.1. Surveillance and Security in Smart Home ................ 22 107 6.1.1. Characteristics of Smart Home for AI/ML processing 108 . 22 109 6.1.2. Use case ......................................... 23 110 6.2. Smart Building Management 111 .............................. 23 112 6.2.1. Characteristics of Smart Building for AI/ML processing23 113 6.2.2. Use case ......................................... 24 114 7. IANA Considerations ........................................ 24 115 8. Acknowledgements ........................................... 24 116 9. Contributors ............................................... 24 117 10. Informative References 118 ..................................... 24 120 1. Introduction 122 The document explains the effects of applying artificial intelligence 123 /machine learning (AI/ML) algorithms in the Internet of Thing (IoT) s 124 ervice environments. 126 IoT applications will be deployed in heterogeneous and different area 127 s such as the energy, transportation, automation and manufacturing in 128 dustries as well as the information and communication technology (IC 129 T) industry. Many IoT sensors and devices can connect to an IoT servi 130 ce environment where IoT objects cannot interoperate with each other 131 and can interact with different applications. The IoT service may not 132 run in a single administrative domain. If market demand exists, the 133 cross-domain service scenarios for IoT applications could be widely d 134 eployed. Future IoT applications occur at multiple domains of heterog 135 eneity with various time scales. 137 The IoT service requirements for common architectures and public APIs 138 poses some challenges to the underlying service environment and netw 139 orking technologies. Some IoT applications require significant securi 140 ty and privacy as well as significant resource and time constraints. 141 These mission-critical applications can be separated from many common 142 IoT applications that current technology may not provide. It means t 144 hat IoT service requirements are difficult to classify common require 145 ments and functional requirements depending on IoT service scenario. 147 Recently, artificial intelligence technologies can help the context-a 148 ware IoT service scenarios apply rule-based knowledge accumulation. T 149 he IoT service assumes that many sensing devices are connected to sin 150 gle or multiple IoT network domains. Each sensor sends small packets 151 to the IoT servers periodically or non-periodically. Detection data c 152 ontains periodic status information that monitors whether the system 153 is in a normal state or not. In some cases, alert information is incl 154 uded for quick processing. Most IoT applications can operate in two m 155 odes. One is a simple monitoring mode and the other is an abnormal mo 156 de for rapid processing. In a simple monitoring phase, the IoT device 157 periodically sends sensing data to the server. If the measured data 158 is outside the normal range, the IoT service can change the operating 159 mode to an abnormal phase and activate future probes. Alarm conditio 160 ns should be promptly notified to responsible persons. For mission-cr 161 itical applications, reliable communication with robust QoS requireme 162 nts in terms of error and latency is required. 164 Periodic data accumulation from IoT devices is cumbersome. Under norm 165 al conditions, the IoT data is simply accumulated without further act 166 ion. In an unusual situation, incoming IoT data can cause an urgent a 167 ction to notify the administrator of the problem. Streaming data traf 168 fic from thousands of IoT devices is annoying to store in the databas 169 e because it is not easy to extract unidentified or future incidents. 170 Only a significant portion of the incoming data stream can be stored 171 in a real-time database that is time-sensitive and capable of rapid 172 query processing. A combination of different IoT detection data, incl 173 uding location, time, and status, allows you to sort and categorize a 174 portion of streaming data when an additional inspection is required, 175 and perform real-time processing. One of the missions of the IoT dat 176 abase is to be able to extract preliminary symptoms of unexpected acc 177 idents from a large amount of streaming data. 179 If some transmitted data is important to invoke the corresponding act 180 ion, there are some questions about whether the incoming data is corr 181 ect. If the incoming data contains accurate and time-critical events, 182 appropriate real-time control and management can be performed. Howev 184 er, if the incoming data is inaccurate or intentionally corrupted, ad 185 ditional accidents may occur. In these cases, incoming data can trigg 186 er to initiate additional inspections to protect against future unacc 187 eptable situations. But, if time-critical data is missed due to error 188 s in the sensing devices and the delivery protocol, there is no reaso 189 n to configure IoT networks and devices at a high cost. 191 It is not easy to analyze data collected through IoT devices installe 192 d to monitor complex IoT service environments. If the sensor malfunct 193 ions, the data of the sensor cannot be trusted. Additional investigat 194 ion should be done if abnormal status from specific sensors is collec 195 ted. The data of the redundant sensor installed in the same area shou 196 ld be received or combined with other sensor information adjacent to 197 the sensor to determine the abnormal state. 199 For sensors installed in a specific area, sensing records will remain 200 for a certain period of time. IoT service operators can look at the 201 operational history of the sensor for a period of time to determine w 202 hat problems were encountered when data was collected. When an abnorm 203 al situation occurs, IoT sensor should investigate whether it noticed 204 normal operations and notified the IoT service operator. If the abno 205 rmal situation is not properly detected, the operator should analyze 206 whether it was caused by malfunction of the IoT sensor or other reaso 207 ns. 209 In the IoT service environment, it is possible to analyze the situati 210 on accurately by applying recent artificial intelligence and machine 211 learning technologies. If there is an operational record of the past, 212 it is possible to determine when an abnormal situation arises. Most 213 problems are likely to be repeated, so if the past learning experienc 214 e is accumulated, the anomaly of IoT services can be easily and immed 215 iately identified. In addition, when information gathered from variou 216 s sensors is synthesized, it is possible to accurately determine whet 217 her abnormal situations have occurred. 219 Various types of IoT sensors are installed with certain purposes. It 220 expects that all the IoT sensors intend to monitor the occurrence of 221 special abnormal situations in advance. Therefore, it should be set i 222 n advance what actions are required when a specific anomaly occurs. T 223 he appropriate work is performed on the abnormal situation according 224 to the procedure, predefined by the human. By using artificial intell 225 igence and machine learning algorithms, the appropriate actions are t 226 aken when an abnormal situation is detected from various IoT sensors. 228 2. Challenging Issues of IoT network 230 This section describes the challenging issues of data sensing, 231 collection, transfer, and intelligent decision from untrusted data 232 quality and unexpected situations of IoT service environments. 234 2.1. Untrusted and incorrect IoT devices 236 IoT traffic is similar to traditional Internet traffic with small 237 packet sizes. Mobile IoT traffic can cause some errors and delays 238 because wireless links are unstable and signal strength may be 239 degraded with device mobility. If the signal strength of the IoT 240 device with a power limit is not so strong, the reception quality of 241 the IoT server may not be sufficient to obtain the measurement data. 243 For mission-critical applications, such as smart-grid and factory- 244 automation, expensive IoT sensors with self-rechargeable batteries 245 and redundant hardware logic may be required. However, unexpected 246 abnormal situations may occur due to sensor malfunctions. There are 247 trade-offs between implementation cost and efficiency for cost- 248 effective IoT services. When smart-grid and factory-automation 249 applications are equipped with IoT devices, the acceptable quality 250 from IoT solutions can be required. Sometimes, expensive and 251 duplicated IoT solutions may be needed. 253 2.2. Traffic burstiness of IoT network 255 IoT traffic includes two types of traffic characteristic: periodic 256 with small packet sizes and bursty with high bandwidth. Under normal 257 conditions, the IoT traffic periodically transmits status information 258 with a small bandwidth, several kilobits/sec. However, in an abnormal 259 state, IoT devices need a high bandwidth, up to several tens of 260 megabits/sec, in order to identify actual events and investigate 261 accurate status information. In addition, traffic volume can 262 explosively increase in response to emergencies. For example, in the 263 case of smart-grid application, the bandwidth of several kilobits/sec 264 is usually used, and when an urgent situation occurs, a broadband 265 channel is required up to several tens of megabits/sec. 267 The other traffic can be integrated at an IoT network to increase 268 bandwidth efficiency. If an emergency situation occurs in the IoT 269 service, IoT traffic volumes suddenly increase, in which case network 270 processing capacity may be not sufficient. If the IoT service is 271 integrated with voice and video applications, the problem can become 272 more complex. As time goes by, traffic congestion and bottlenecks are 273 frequent in some areas. In addition, if an existing service policy 274 changes (for example, prioritizing certain traffic or suddenly 275 changing the route), other unexpected problems may be encountered. 276 Various congestion control and load balancing algorithms with the 277 help of artificial intelligence can be applied to handle time-varying 278 traffic on a network. 280 Until now, much research has been done on traffic variability in an 281 integrated network service environment. All networks have their own 282 traffic characteristics, depending on geographical area, number of 283 subscribers, subscribers' preferences, and types of applications used. 284 In the case of IoT traffic, the normal bandwidth is very small. If 285 the IoT traffic volume increases abruptly in an abnormal situation, 286 the network may suffer unacceptable delay and loss. If emergency 287 situations detected by IoT networks occur in a smart grid or 288 intelligent transportation system, the processing power of the IoT 289 network alone cannot solve the problem and the help of existing 290 network resources is inevitable. 292 2.3. Management overheads of heterogeneous IoT sensors 294 Traffic management in an integrated network environment is not easy. 295 In order to operate the network steadily, a network operator has its 296 own know-hows and experiences. If there are plenty of network 297 resources, it is easy to set up a bypass route even if network 298 failure or congestion occurs in a specific area. For operating 299 network steadily, network resources may be designed to be over- 300 provisioned in order to cope with various possible outages. A network 301 operator predicts the amount of traffic generated by the 302 corresponding equipment and grasps to what extent a transmission 303 bandwidth is required. If traffic fluctuation is very severe, the 304 network operator can allocate network resources in advance. In case 305 of frequent failures or severe traffic fluctuation, some network 306 resources are separated in order not to affect normal traffic. 308 More than a billion IoT devices are expected to connect to 309 smartphones, tablets, wearables, and vehicles. Therefore, IoT 310 services are targeted at mobile applications. In particular, 311 intelligent transportation systems need the help of IoT technology to 312 provide traffic monitoring and prevent public or private traffic 313 accidents. IoT technology can play an important role in reducing 314 traffic congestion, saving people's travel time and costs, and 315 providing a pleasant journey. 317 The IoT service has troublesome administrative problems to configure 318 an IoT network which consists of IoT servers, gateways, and many 319 sensing devices. The small-sized but large-numbered IoT devices may 320 incur administrative overhead since all the IoT devices should be 321 initialized and the bootstrapping information of IoT resources should 322 be loaded into the IoT service environments. Whenever some IoT 323 devices are newly added and some devices have to be removed, the 324 dynamic reconfiguration of IoT resources is essential. In addition, 325 the IoT device's preinstalled software should be regularly inspected 326 and upgraded according to its version. Frequent upgrades and changes 327 to some IoT devices may require autonomic management and 328 bootstrapping techniques. 330 Network management generally assumes that all network resources 331 operate reliably with acceptable quality. In most failure situations, 332 the network operator decides to switch to a redundant backup device 333 or bypass the failed communication path. If some IoT devices are not 334 stable, duplicate IoT devices can be installed for the same purpose. 335 If IoT resources are not duplicated, various mechanisms are needed to 336 reduce the damage. Therefore, it is necessary to prioritize the 337 management tasks to be performed first when an abnormality occurs in 338 the IoT service environment. However, managing duplicate networks can 339 cause another problem. If two IoT devices are running at the same 340 time, the recipient can get redundant information. If two or more 341 unusual situations occur at the same time, it is difficult to solve 342 the problem since tasks for urgent processing should be distinguished 343 from tasks that can be performed over time. 345 In addition, the operations manager's mistakes or misunderstanding of 346 problem situations can lead to other unexpected complications. 347 Therefore, artificial intelligence technologies can help what kind of 348 network management work is required when an unexpected complicated 349 situation occurs even though a procedure for an abnormal situation is 350 already prepared. 352 3. Overview of AI/ML-based IoT services 354 In this section, successful applications of artificial intelligence i 355 n IoT domains are provided. The common property of IoT applications a 356 nd services is that they require fast analytics rather than later ana 357 lytics with piled data. Recently, neural-network-based artificial int 358 elligence technologies are widely used across many IoT applications. 360 Simple IoT applications include dynamic contexts that share common fe 361 atures among social relations at the same administration domain. IoT 362 devices in the same domain can provide their service contexts to the 363 IoT server. When a dynamic change occurs in an IoT service context, t 364 he IoT device needs real-time processing to activate urgent events, a 365 lert notifications, update, and reconnect contexts. The IoT service m 366 ust support real-time interactions between the IoT device and the sys 367 tem in the same domain. The IoT service contexts must be shared betwe 368 en physical objects and social members in the same domain as well. 370 Artificial intelligence technologies have been shown promising in man 371 y areas, including IoT. For example, contextual information for a car 372 -sharing business must interact with customers, car owners, and car s 373 haring providers. All entities in the value chain of a car sharing bu 374 siness must share the corresponding situation to pick up, board, and 375 return shared cars. Communication networks and interactive informatio 376 n, including registration and payment, can be shared tightly among th 377 e entities. Home IoT service environment can be equipped with sensors 378 for theft detection, door lock, temperature, fire detection, gas det 379 ection, short circuit, air condition to name a few. Office IoT servic 380 e environments, including buildings such as shopping centers and bus/ 381 airport terminals, have their own sensors, including alarm sensors. W 382 hen an alarm signal is detected by the sensor, the physical position 383 and occurrence time of the sensor is determined in advance. All signa 384 ls from various sensors are analyzed comprehensively to make the righ 385 t decision. If some sensors frequently malfunction, the situation can 386 be grasped more accurately by analyzing the information of the adjac 387 ent sensor. In particular, when installing multiple sensors in a part 388 icular building (e.g., surveillance camera, location monitoring, temp 389 erature, etc.), a much wider range of sensors can be used when utiliz 390 ing artificial intelligence and machine learning technologies. 392 (Smart home) Smart home concept span over multiple IoT applications, 393 health, energy, entertainment, education, etc. It involves voice reco 394 gnition, natural language processing, image-based object recognition, 395 appliance management, and many more artificial intelligence technolo 396 gies integrated with IoT. Smart connected-devices monitor the house t 397 o provide better control over home supplies and expenses. The energy 398 consumption and efficiency of home appliances are monitored and analy 399 zed with deep learning based technologies, such as artificial neural 400 network, long-short-term-memory, etc. 402 (Smart city) Smart city, as well, contains multiple IoT domains, tran 403 sportation, infrastructure, energy, agriculture, etc. Since heterogen 404 eous data from different domains are gathered in smart cities, variou 405 s artificial intelligence approaches are studied in smart-city applic 406 ation. Public transportation behaviors and crowd movements patterns a 407 re important issues, and they are often dealt with neural network bas 408 ed methods, long-short-term-memory and convolutional neural network. 410 (Smart energy) As two-way communication energy infrastructure is depl 411 oyed, smart grid has become a big IoT application, which requires int 412 elligent data processing. The traditional energy providers are highly 413 interested in recognizing local energy consumption patterns and fore 414 casting the needs in order to make appropriate decisions on real-time. 415 Moreover, the energy consumers, as well, want analyzed information o 416 n their own energy consumption behaviors. Recently, many works on ene 417 rgy consumption prediction, energy flexibility analysis, etc. are act 418 ively ongoing. Most works are based on the latest deep learning techn 419 ologies, such as multi-layered-perceptron, recurrent neural network, 420 long-short-term-memory, autoencoder, etc. 422 (Smart transportation) The intelligent transportation system is anoth 423 er source of big data in IoT domains. Many use cases, such as traffic 424 flow and congestion prediction, traffic sign recognition, vehicle in 425 trusion detection, etc., have been studied. Moreover, a lot of advanc 426 ed artificial intelligence technologies are required in autonomous an 427 d smart vehicles, which require many intelligent sub-tasks, such as p 428 edestrian's detection, obstacle avoidance, etc. 430 (Smart healthcare) IoT and artificial intelligence are integrated int 431 o the healthcare and wellbeing domain as well. By analyzing food imag 432 es with convolutional neural network on mobile devices, dietary intak 433 es can be measured. With voice signal captured from sensor devices, v 434 oice pathologies can be detected. Moreover, recurrent neural network 435 and long-short-term-memory technologies are actively being studied fo 436 r early diagnosis and prediction of diseases with time series medical 437 data. 439 (Smart agriculture) To manage a vast area of land, IoT and artificial 440 intelligence technologies are recently used in agriculture domains. 441 Deep neural network and convolutional neural network are utilized for 442 crop detection or classification and disease recognition in the plan 443 ts. Moreover, for automatic farming with autonomous machine operation, 444 obstacle avoidance, fruit location, and many more sub-tasks are hand 445 led with advanced artificial intelligence technologies. 447 4. Requirements for AI/ML-based IoT services 449 In this section, the requirements for AI/ML-based IoT data collection 450 and delivery, intelligent and context-aware IoT services, and applyi 451 ng AI/ML to IoT data will be described. 453 4.1. Requirements for AI/ML-based IoT data collection and delivery 455 IoT services store a vast amount of data that IoT devices 456 periodically generate, and the refining and analyzing are costly. 457 Effective analysis of IoT data has been considered to be the most 458 important factor in data processing, but the analysis of efficient 459 data collection and delivery methods are becoming other significant 460 factors as the amount of the data collected is explosively increasing. 462 In particular, as a number of IoT devices have been deployed within 463 the IoT network, controlling data collection and delivery for each of 464 them has become impossible. The introduction of AI/ML techniques for 465 simultaneous and efficient management of the IoT devices should be 466 considered as a countermeasure. For IoT data collection and delivery, 467 the following two factors will need to be considered, IoT devices 468 energy and data quality. 470 (IoT Device Energy) As many IoT devices have begun to be deployed 471 within the IoT network, it is impossible to deliver energy to many 472 IoT devices simultaneously. Consequently, the efficient battery use 473 has become an important issue. 475 If IoT data collection and delivery periods are too short, a lifetime 476 of the IoT device will be shortened through the reckless use of IoT 477 device energy. Thereby, it increases the cost required to provide IoT 478 service. On the other hand, if IoT data collection and delivery 479 period are too long, the quality of the IoT services provided will be 480 reduced due to the lack of details in the data for situation 481 recognition and real-time processing. Therefore, taking into account 482 the energy consumption of the IoT devices, research on proper IoT 483 data collection and delivery period is necessary. 485 (Data Quality) Since the data collected from the majority of IoT 486 devices usually contain redundant information, it causes additional 487 costs for the data collection and refinement processes. Therefore, it 488 will be necessary to select and deliver meaningful information from 489 redundant IoT data to reduce unnecessary cost on the IoT network. To 490 do so, it will need the research to identify the relationships among 491 the data collected various devices and interpret the information that 492 the data contains. 494 4.2. Requirements for intelligent and context-aware IoT services 496 In a context-aware IoT service environment, it is important to 497 establish a context to be aware of in advance since IoT devices will 498 be deployed according to a pre-designed architecture and to check how 499 characteristics of IoT data and data-to-data characteristics are 500 expressed under these circumstances. For the data produced by IoT 501 devices, since it contains the device's relative location information, 502 sensing value over time, event information, it should be reviewed to 503 provide the target context-aware service using this information. Some 504 of the necessary technologies will be described in the following. 506 (Physical Clustering) To increase the accuracy of context-awareness, 507 the provision of context-aware services should be considered in a 508 situation where the relationship between IoT devices with respect to 509 physical layout or physical environment is taken into account. 510 Setting a rule using the service provider's domain knowledge may be 511 possible, but introducing the physical clustering into a diverse IoT 512 environment (e.g., in bedroom, kitchen, balcony, or a space connected 513 through an open door) will require identifying the physical 514 relationship between the devices using data generated from IoT 515 devices. 517 (Extra Data Processing) In order to prevent degradation of service 518 quality from errors in data values or device malfunctions, extra 519 sensors should be placed in the majority of IoT environments. In a 520 context-aware service, they contain the same information, so the 521 technologies filtering the data that contains only essential part 522 among the same information while preventing data errors would be 523 required. 525 (Unreported data handling) If an event is detected on a particular 526 IoT device, it will transmit data regardless of the device's sensing 527 and delivery interval. At this time, the data of IoT devices which 528 are physically clustered are needed to accurately detect events that 529 occurred, and it is difficult to expect that these devices will 530 provide data at the same time. A gateway can request data from 531 clustered devices, but it has a problem for real-time processing for 532 emergency situations. Therefore, handling unreported data will be 533 required based on previously collected data. 535 (Abnormal data in AI/ML) In the case of context-aware services that 536 operates based on the predetermined rule, the flexibility to cope 537 with emergency situations that have not been considered is low, and 538 thus AI/ML algorithms are required to intelligently cope with a 539 myriad of situations. However, many abnormal data are generated 540 depending on environmental conditions such as device status, so AI/ML 541 algorithms that can operate in that imperfect environment should be 542 considered. 544 (Edge computing in IoT) There are two necessary prerequisites 545 required in context-aware IoT services: IoT devices real-time 546 management and IoT network architecture supporting the high-volume 547 data transmission. When an abnormal situation is discovered, high- 548 volume data should be utilized to adequately to monitor the situation 549 through the IoT device's real-time management. As contrasted with 550 conventional cloud computing structures, an edge computing structure, 551 where IoT data processing servers are located in closer proximity to 552 IoT devices, provides higher energy efficiency and lesser 553 communication delay. 555 4.3. Requirements for applying AI/ML to IoT data 557 In this subsection, the requirements for applying AI/ML to IoT data 558 are described. 560 4.3.1. Training AI/ML algorithm 562 To use AI/ML algorithm, two elements are required, AI/ML model and 563 training data. The presence of training dataset in good quality is an 564 important factor of the AI/ML model performance since the model is 565 iteratively trained with the training data. However, for anomaly 566 detection, there is not enough training data since not only the 567 probability of anomaly occurrence is very low but also it is almost 568 impossible to retrieve the ground truth value even when the situation 569 has occurred. Therefore, using domain knowledge, AI/ML learning based 570 on abnormal situation data generation or simulation should be 571 considered. For example, for an external intrusion detection 572 application within a smart home, when a camera and a motion sensor 573 detect an intruder, a light sensor checks the measuring value. If the 574 light does not turn on, then the IoT application recognizes it as an 575 abnormal situation. In this way, by using the domain knowledge, the 576 rule regarding the operational scenario of the IoT application is 577 generated as the training data, and the generated training data can 578 be used for model learning. This will not only enable learning the 579 anomaly detection algorithm in IoT application but also improving the 580 accuracy. Therefore, IoT application, in which it is difficult to 581 acquire dataset in good quality, will require data generation based 582 on domain knowledge for AI/ML. 584 4.3.2. AI/ML inference in IoT application 586 In order for AI ML technology to be applied to IoT applications, the 587 training data and the input data for model testing and inferencing 588 must have the same characteristics such as dimension, time interval, 589 types of features, etc. However, due to the volatile IoT data 590 characteristics that vary from situations in many IoT applications, 591 it is difficult to directly apply the AI/ML algorithms. For example, 592 in a simple monitoring phase, the IoT devices periodically send 593 sensing data, and AI/ML have no difficulty in operating. However, in 594 an abnormal mode, the IoT applications require a fast response, and 595 IoT devices transmit data at shorter intervals than normal, which 596 changes the characteristics of the data being input to the AI/ML 597 algorithm. Therefore, data preprocessing technology handling the 598 abnormal data will be required in advance, such as data imputation, 599 correcting data anomalies, and Interpolation of unreported data. 601 5. State of arts of the artificial intelligence/machine learning 602 technologies for IoT services 604 In this section, well-known machine learning and artificial 605 intelligence technologies applicable to IoT applications are reviewed. 607 5.1. Machine learning and artificial intelligence technologies review 609 The classical machine learning models can be divided into three types, 610 supervised, unsupervised, and reinforcement learnings. Therefore, in 611 this subsection, machine learning and artificial intelligence 612 technology reviews are done in four different categories: supervised, 613 unsupervised, reinforcement, and neural-network-based. 615 5.1.1. Supervised learning for IoT 617 Supervised learning is a task-based type of machine learning, which 618 approximates function describing the relationship and causality 619 between input and output data. Therefore, the input data needs to be 620 clearly defined with proper output data since supervised learning 621 models learn explicitly from direct feedback. 623 (K-Nearest Neighbor) Given a new data point in K-Nearest Neighbor 624 (KNN) classifier, it is classified according to its K number of the 625 closest data points in the training set. To find the K nearest 626 neighbors of the new data point, it needs to use a distance metric 627 which can affect classifier performance, such as Euclidean, 628 Mahalanobis or Hamming. One limitation of KNN in applying for IoT 629 network is that it is unscalable to large datasets because it 630 requires the entire training dataset to classify a newly incoming 631 data. However, KNN required less processing power capability compared 632 to other complex learning methods. 634 (Naive Bayes) Given a new data point in Naive Bayes classifiers, 635 it is classified based on Bayes' theorem with the "naive" assumption 636 of independence between the features. Since Naive Bayes classifiers 637 don't need a large number of data points to be trained, they can deal 638 with high-dimensional data points. Therefore, they are fast and 639 highly scalable. However, since its "naive" assumptions are somewhat 640 strong, a certain level of prior knowledge on the dataset is required. 642 (Support Vector Machine) Support Vector Machine (SVM) is a binary and 643 non-probabilistic classifier which finds the hyperplane maximizing 644 the margin between the classes of the training dataset. SVM has been 645 the most pervasive machine learning technology until the study on 646 neural network technologies are advanced recently. However, SVM still 647 has advantages over neural network based and probabilistic approaches 648 in terms of memory usage and capability to deal with high-dimensional 649 data. In this manner, SVM can be used for IoT applications with 650 severe data storage constraint. 652 (Regression) Regression is a method for approximating the 653 relationships of the dependent variable, which is being estimated, 654 with the independent variables, which are used for the estimation. 655 Therefore, this method is widely used for forecasting and inferring 656 causal relationships between input data and output data in time- 657 sensitive IoT application. 659 (Random Forests) In random forests, instead of training a single 660 decision tree, a group of trees is trained. Each tree is trained on a 661 subset of the training set using a randomly chosen subset of M input 662 variables. Random forests considering various tree structures have 663 very high accuracy, so it can be utilized in the accuracy-critical 664 IoT applications. 666 5.1.2. Unsupervised learning for IoT 668 Unsupervised learning is a data-driven type of machine learning which 669 finds hidden structure in unlabeled dataset without feedback during 670 the learning process. Unlike supervised learning, unsupervised 671 learning focuses on discovering patterns in the data distributions 672 and gaining insights from them. 674 (K-means clustering) K-means clustering aims to assign observations 675 into K number of clusters in which each observation belongs to the 676 cluster having the most similarities. The measure of similarity is 677 the distance between K cluster centers and each observation. K-means 678 is a very fast and highly scalable clustering algorithm, so it can be 679 used for IoT applications with real-time processing requirements such 680 as smart transportation. 682 (Density-based spatial clustering of applications with noise) 683 Density-Based approach to Spatial Clustering of Applications with 684 Noise (DBSCAN) is a method that clusters dataset based on the density 685 of its data samples. In this model, dense regions which include data 686 samples with many close neighbors are considered as clusters, and 687 data samples in low-density regions are classified as outliers 688 [Kriegal]. Since this method is robust to outliers, DBSCAN is 689 efficient data clustering method for IoT network environments with 690 untrusted big datasets in practice. 692 5.1.3. Reinforcement learning for IoT 694 Reinforcement learning is a reactive type of machine learning that 695 learn a series of actions in a given set of possible states, actions, 696 and rewards or penalties. It can be seen as the exploring decision- 697 making process and choosing the action series with the most reward or 698 the least penalty which can be cost, priority, time to name a few. 699 Reinforcement learning can be helpful for selecting action of IoT 700 device by providing a guideline. 702 (Q-learning) Q-Learning is a model-free, off-policy reinforcement 703 learning algorithm based on the well-known Bellman Equation. The goal 704 is to learn an action-selection policy maximizing the Q-value, which 705 tells an agent what action to take. It can be used for IoT device to 706 determine which action it should take according to conditions. 708 (State-Action-Reward-State-Action) Though State-Action-Reward-State- 709 Action (SARSA) is a much similar algorithm to Q-learning, the main 710 difference is that it is an on-policy algorithm in which agent 711 interacts with the environment and updates the policy based on 712 actions taken. It means that the Q-value is updated by an action 713 performed by the current policy instead of the greed policy that 714 maximizes Q-value. In this perspective, it is relevant when an action 715 of one IoT device will greatly influence the condition of the 716 environment. 718 (Deep Q Network) Deep Q network (DQN) is developed to solve the 719 exploration problem for unseen states. In the case of Q-learning, the 720 agent is not capable of estimating value for unseen states. To handle 721 this generality problem, DQN leverages neural network technology. As 722 a variation of the classic Q-Learning algorithm, DQN utilizes a deep 723 convolutional neural net architecture for Q-function approximation. 724 In real environments not all possible states and conditions are not 725 able to be observed. Therefore, DQN is more relevant than Q-learning 726 or SARSA in real applications such as IoT. Since DQN could be used 727 within only discrete action space, it can be utilized for traffic 728 routing in the IoT network. 730 (Deep Deterministic Policy Gradient) DQN has solved generality and 731 exploration problem of the unseen or rare states. Deep Deterministic 732 Policy Gradient (DDPG) takes DQN into the continuous action domain. 733 DDPG is a deterministic policy gradient based actor-critic, model- 734 free algorithm. The actor decides the best action for each state and 735 critic is used to evaluate the policy, the chosen action set. In IoT 736 applications, DDPG can be utilized for the tasks that require 737 controlled in continuous action spaces, such as energy-efficient 738 temperature control, computation offloading, network traffic 739 scheduling, etc. 741 5.1.4. Neural Network based algorithms for IoT 743 (Recurrent Neural Network) Recurrent Neural Network (RNN) is a 744 discriminative type of supervised learning model that takes serial or 745 time-series input data. RNN is specifically developed to address 746 issue of time dependency of sequential time-series input data. It 747 processes sequences of data through internal memory, and it is useful 748 in IoT applications with time-dependent data, such as identifying 749 time-dependent patterns of sensor data, estimating consumption 750 behavior over time, etc. 752 (Long Short Term Memory) As an extension of RNN, Long Short Term 753 Memory (LSTM) is a discriminative type of supervised learning model 754 that is specialized for serial or time-series input data as well 755 [Hochreiter]. The main difference of LSTM from RNN is that it 756 utilizes the concept of gates. It actively controls forget gates to 757 prevent the long term time dependency from waning. Therefore, 758 compared to RNN, it is more suitable for data with long time 759 relationship and IoT applications requiring analysis on the long lag 760 of dependency, such as activity recognition, disaster prediction, to 761 name a few [Chung]. 763 (Convolutional Neural Network) Convolutional neural network (CNN) is 764 a discriminative type of supervised learning model. It is developed 765 specifically for processing 2-dimensional image data by considering 766 local connectivity, but now generally used for multidimensional data 767 such as multi-channel sound signals, IoT sensor values, etc. As in 768 CNN neurons are connected only to a small subset of the input and 769 share weight parameters, CNN is much more sparse compared to fully 770 connected network. However, it needs a large training dataset, 771 especially for visual tasks. In CNN, a new activation function for 772 neural network, Rectified Linear Unit (ReLU), was proposed, which 773 accelerates training time without affecting the generalization of the 774 network [Krizhevsky]. In IoT domains, it is often used for detection 775 tasks that require some visual analysis. 777 (Variational Autoencoder) Autoencoder (AE) is a generative type 778 of unsupervised learning model. AE is trained to generate output to 779 reconstruct input data, thus it has the same number of input and 780 output units. It is suitable for feature extraction and 781 dimensionality reduction. Because of its behavior to reconstructing 782 the input data at the output layer, it is often used for machinery 783 fault diagnosis in IoT applications. The most popular type of AE, 784 Variational Autoencoder (VAE) is a generative type of semi-supervised 785 learning model. Its assumptions on the structure of the data are weak 786 enough for real applications and its training process through 787 backpropagation is fast [Doersch]. Therefore, VAE is suitable in IoT 788 applications where data tends to be diverse and scarce. 790 (Generative Adversarial Network) Generative Adversarial Network (GAN) 791 is a hybrid type of semi-supervised learning model which contain two 792 neural networks, namely the generative and discriminative networks 793 [Goodfellow]. The generator is trained to learn the data distribution 794 from a training dataset in order to generate new data which can 795 deceive the latter network, so-called the discriminator. Then, the 796 discriminator learns to discriminate the generated data from the real 797 data. In IoT applications, GAN can be used in situations when 798 something needs to be generated from the available data, such as 799 localization, way-finding, and data type conversion. 801 5.2. Technologies for lightweight and real-time intelligence 803 As the era of IoT has come, some sort of light-weight intelligence is 804 needed to support smart objects. Prior to the era of IoT, most of the 805 works on learning did not consider resource-constrained environments. 806 Especially, deep learning models require many resources such as 807 processing power, memory, stable power source, etc. However, it has 808 been recently shown that the parameters of the deep learning models 809 contain redundant information, so that some parts of them can be 810 delicately removed to reduce complexity without much degradation of 811 performance [Ba], [Denil]. In this section, the technologies to 812 achieve real-time and serverless learning in IoT environments are 813 introduced. 815 (network compression) Network compression is a method to convert a 816 dense network into a sparse one. With this technology the network can 817 be reduced in its size and complexity. By pruning irrelevant parts or 818 sharing redundant parameters, the storage and computational 819 requirements can be decreased [Han]. After pruning, the performance 820 of the network is examined and the pruning process is repeated until 821 the performance reaches the minimum requirements for the specific 822 applications and use cases. As many parameters are removed or shared, 823 the memory required is reduced, as well as computational burden and 824 energy. Especially as most energy in neural network is used to access 825 memory, the consumed energy dramatically drops. Although its main 826 limitation is that there is not a general solution to compress all 827 kinds of network, but it rather depends on the characteristics of 828 each network. However, network compression is still the most 829 widespread method to make deep learning technologies to be 830 lightweight and IoT-friendly. 832 (approximate computing) Approximate computing is an approach to 833 support deep learning in smart devices [Venkataramani], [Moons]. It 834 is based on the facts that the results of deep learning do not need 835 to be exact in many IoT applications but still valid if the results 836 are in an acceptable range. By integrating approximate computing into 837 deep learning, not only the execution time but also the energy 838 consumption is reduced [Mohammadi]. Based on the optimal trade-off 839 between accuracy and run-time or energy consumption, the network can 840 be adjustably approximated. The network approximate technology can be 841 well-used in such situations when the response time is more important 842 than sophisticatedly analyzed results. Although it is a technology to 843 facilitate real-time and lightweight intelligence, the process of 844 training models and converting it to approximate network require some 845 amount of resource. Therefore, the approximated model can be deployed 846 on smart devices but the learning and approximation processes still 847 need to take places on resource rich platforms. 849 6. Use cases of AI/ML into IoT service 851 Many IoT service environments are equipped with camera, door lock, 852 temperature sensor, fire detector, gas detector, alarm, and so on. 853 Each sensor is deployed with particular purposes of each own to 854 provide a specific service. However, there is a problem that the 855 sensor utilization is not high enough due to the provision of the 856 service using only a single sensor rather than multiple sensors and 857 their mutual relations. Therefore, the quality of the service 858 provided is not high as well. To enhance the sensor utilization and 859 the service quality, all signals from various sensors should be 860 analyzed comprehensively to make the right decision. This section 861 describes the use cases for introducing AI / ML techniques in actual 862 IoT service, utilizing multiple sensors. In advance of each use case 863 description of various IoT service domains, characteristics of each 864 domain to adopt AI/ML techniques are investigated. 866 6.1. Surveillance and Security in Smart Home 868 To minimize users?inconvenience and ensure their safety, 869 surveillance and safety IoT applications provided within smart homes 870 require fast notification with good level of precision IoT service 871 quality for abnormal conditions detection. To do this, both data 872 preprocessing techniques and AI/ML technologies for analysis of 873 anomalies with high accuracy will be required. 875 6.1.1. Characteristics of Smart Home for AI/ML processing 877 (Training Data Generation) For Surveillance and Security, the 878 processed data is necessary because there is little data for 879 anomalies and the ground truth values are hardly available. Therefore, 880 first, the steps to detect and calibrate the abnormal data are 881 essential before the anomaly data should be generated using domain 882 knowledge. First, constructing simulators about targeted smart home 883 and generating events against external intrusions and then collecting 884 the anomaly data can be considered. Furthermore, based on the data 885 collected in the actual environment, anomaly data generation can 886 proceed by breaking the relationship between sensors considering 887 possible links between them within any intrusive environment. 889 (AI/ML Algorithm) One of the characteristics of the IoT environment 890 for surveillance and safety is that a massive amount of data is 891 collected and real-time responses are required. For the kNN algorithm, 892 since the more data sets, the stronger against the noise and the 893 higher the accuracy. If the appropriate dataset is used, the fast 894 response can be expected. It makes suitable for the service 895 environment to be considered. In addition, considering real-time data 896 forecasting and analysis via LSTM, it is believed that improved 897 accuracy for real-time anomalies detection can be expected. 899 6.1.2. Use case 901 (To be continued) 903 6.2. Smart Building Management 905 Smart buildings often consist of heterogeneous IoT devices. These 906 devices cooperate and their data is integrated for efficient 907 autonomous building management. Many of the events in a large 908 building may not require deep, complicated learning or processing. 909 Some of them may require a fast response than an accurate analysis. 910 Above all, a lot of events simultaneously occur and can arise heavy 911 loads on the main server. The edge-computing techniques can be used 912 to offload the main server's tasks. 914 6.2.1. Characteristics of Smart Building for AI/ML processing 916 (Training Data Generation) In smart buildings, heterogeneous IoT 917 devices are deployed. They are diverse in their types, functions, 918 performances, etc. To utilize the data from diverse devices, data 919 needs to be able to well-integrated. Therefore, it is better for data 920 to be in a common data format, or it needs to be able to transform 921 into one another. The other characteristic is that the IoT devices 922 may interact in local and global environments of the building. 923 Therefore, the scope of the dataset used in training can play a 924 critical role in developing AL/ML model for building management. 926 (AI/ML Algorithm) To offload and reduce the burden of the main server 927 and to provide fast, efficient decision makings, the IoT and the 928 other network-related devices can use their computing resources. 929 Various edge-computing techniques can be applied to do so, such as 930 developing light-weighted AI/ML models that can be easily deployed in 931 the edge devices or balancing the learning and processing computation 932 load from the server to the edge devices. 934 6.2.2. Use case 936 (To be continued) 938 7. IANA Considerations 940 This document requests no action by IANA. 942 8. Acknowledgements 944 9. Contributors 946 10. Informative References 948 [Hochreiter]S. Hochreiter and J. Schmidhuber, "Long short-term 949 memory," Neural Comput., vol. 9, no. 8, pp. 1735-1780, Nov. 950 1997. 952 [Chung] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, "Empirical 953 evaluation of gated recurrent neural networks on sequence 954 modeling," arXiv preprint arXiv:1412.3555v1 [cs.NE], 2014. 956 [Krizhevsky]A. Krizhevsky, I. Sutskever, and G. E. Hinton, "ImageNet 957 classification with deep convolutional neural networks," in 958 Proc. Adv. Neural Inf. Process. Syst., 2012, pp. 1097-1105.. 959 Hochreiter and J. Schmidhuber, "Long short-term memory," 960 Neural Comput., vol. 9, no. 8, pp. 1735-1780, Nov. 1997. 962 [Doersch] C. Doersch, "Tutorial on variational autoencoders," arXiv 963 preprint arXiv:1606.05908v2 [stat.ML], 2016. 965 [Goodfellow]I. Goodfellow et al., "Generative adversarial nets," in 966 Proc. Adv. Neural Inf. Process. Syst., 2014, pp. 2672-2680. 968 [Ba] J. Ba and R. Caruana, "Do deep nets really need to be deep?" 969 in Proc. Adv. Neural Inf. Process. Syst., Montreal, QC, 970 Canada, 2014, pp. 2654-2662. 972 [Denil] M. Denil, B. Shakibi, L. Dinh, N. de Freitas, and M. Ranzato, 973 "Predicting parameters in deep learning," in Proc. Adv. Neural Inf. 974 Process. Syst., 2013, pp. 2148-2156. 976 [Han] S. Han, J. Pool, J. Tran, and W. Dally, "Learning both 977 weights and connections for efficient neural network," in 978 Proc. Adv. Neural Inf. Process. Syst., Montreal, QC, Canada, 979 2015, pp. 1135-1143. 981 [Venkataramani]S. Venkataramani, A. Ranjan, K. Roy, and A. 982 Raghunathan, "AxNN: Energy-efficient neuromorphic systems 983 using approximate computing," in Proc. Int. Symp. Low Power 984 Electron. Design, ACM, 2014, pp. 27-32. [Moons]S. 985 Hochreiter and J. Schmidhuber, "Long short-term memory," 986 Neural Comput., vol. 9, no. 8, pp. 1735-1780, Nov. 1997. 988 [Moons] B. Moons, B. De Brabandere, L. Van Gool, and M. Verhelst, 989 "Energy- efficient ConvNets through approximate computing," 990 in Proc. IEEE Winter Conf. Appl. Comput. Vis. (WACV), Lake 991 Placid, NY, USA: IEEE, 2016, pp. 1-8. 993 [Mohammadi] Mohammadi, Mehdi, et al. "Deep learning for IoT big data 994 and streaming analytics: A survey," IEEE Communications 995 Surveys & Tutorials, 2018, pp. 2923-2960. 997 [Kriegel] Kriegel, HansPeter, et al. "Densitybased clustering," 998 Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 999 2011, pp. 231-240. 1001 Authors?Addresses 1003 Jun Kyun Choi (editor) 1004 Korea Advanced Institute of Science and Technology (KAIST) 1005 193 Munji Ro, Yuseong-gu, Daejeon, Korea 1007 Email: jkchoi59@kaist.ac.kr 1009 Jae Seob Han 1010 Korea Advanced Institute of Science and Technology (KAIST) 1011 193 Munji Ro, Yuseong-gu, Daejeon, Korea 1013 Email: j89449@kaist.ac.kr 1015 Gyeong Ho Lee 1016 Korea Advanced Institute of Science and Technology (KAIST) 1017 193 Munji Ro, Yuseong-gu, Daejeon, Korea 1019 Email: gyeongho@kaist.ac.kr 1021 Na Kyoung Kim 1022 Korea Advanced Institute of Science and Technology (KAIST) 1023 193 Munji Ro, Yuseong-gu, Daejeon, Korea 1025 Email: nkim71@kaist.ac.kr