idnits 2.17.00 (12 Aug 2021) /tmp/idnits37836/draft-tavernier-irtf-lccn-problem-statement-01.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: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (January 15, 2011) is 4144 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- == Unused Reference: 'AI-modern' is defined on line 661, but no explicit reference was found in the text == Unused Reference: 'PRML' is defined on line 680, but no explicit reference was found in the text == Outdated reference: A later version (-06) exists of draft-ietf-idr-bgp-issues-03 Summary: 0 errors (**), 0 flaws (~~), 4 warnings (==), 1 comment (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Internet Research Task Force W. Tavernier, Ed. 3 Internet-Draft Ghent University - IBBT 4 Intended status: Informational D. Papadimitriou 5 Expires: July 19, 2011 Alcatel-Lucent Bell 6 D. Colle 7 Ghent University - IBBT 8 January 15, 2011 10 Learning Capable Communication Network (LCCN) problem statement 11 draft-tavernier-irtf-lccn-problem-statement-01 13 Abstract 15 Operational procedures and protocols of today's communication 16 networks typically use explicitly defined mechanisms and 17 representations to reach the goals associated to their design. This 18 practice results into numerous protocols having a restricted space 19 for (self-)adaptability, flexibility, and sensitivity respective to 20 their network context (e.g. network traffic conditions, failure 21 conditions, etc). On the other hand, a wide spectrum of learning and 22 optimization techniques is available such that networks could learn 23 and optimize their behavior in the running context. This document 24 describes the opportunities and challenges for a Learning Capable 25 Communication Network (LCCN). 27 Status of this Memo 29 This Internet-Draft is submitted in full conformance with the 30 provisions of BCP 78 and BCP 79. 32 Internet-Drafts are working documents of the Internet Engineering 33 Task Force (IETF). Note that other groups may also distribute 34 working documents as Internet-Drafts. The list of current Internet- 35 Drafts is at http://datatracker.ietf.org/drafts/current/. 37 Internet-Drafts are draft documents valid for a maximum of six months 38 and may be updated, replaced, or obsoleted by other documents at any 39 time. It is inappropriate to use Internet-Drafts as reference 40 material or to cite them other than as "work in progress." 42 This Internet-Draft will expire on July 19, 2011. 44 Copyright Notice 46 Copyright (c) 2011 IETF Trust and the persons identified as the 47 document authors. All rights reserved. 49 This document is subject to BCP 78 and the IETF Trust's Legal 50 Provisions Relating to IETF Documents 51 (http://trustee.ietf.org/license-info) in effect on the date of 52 publication of this document. Please review these documents 53 carefully, as they describe your rights and restrictions with respect 54 to this document. Code Components extracted from this document must 55 include Simplified BSD License text as described in Section 4.e of 56 the Trust Legal Provisions and are provided without warranty as 57 described in the Simplified BSD License. 59 Table of Contents 61 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 3 62 2. Learning opportunities . . . . . . . . . . . . . . . . . . . . 4 63 2.1. Availability of network data and statistics . . . . . . . 4 64 2.2. Availability of processing capacity . . . . . . . . . . . 5 65 3. The learning process . . . . . . . . . . . . . . . . . . . . . 5 66 4. Architectural implications . . . . . . . . . . . . . . . . . . 7 67 4.1. From a pre-defined open-loop control towards a 68 self-adaptive closed-loop control . . . . . . . . . . . . 7 69 4.2. The integration of learning capability . . . . . . . . . . 9 70 4.3. Coexistance with current networking protocols, 71 mechanisms and practices . . . . . . . . . . . . . . . . . 10 72 4.4. Complexity/control vs. performance/labour trade-off 73 measurability . . . . . . . . . . . . . . . . . . . . . . 10 74 5. Applicability . . . . . . . . . . . . . . . . . . . . . . . . 11 75 5.1. Functional domains . . . . . . . . . . . . . . . . . . . . 11 76 5.2. Scope with respect to the hourglass model . . . . . . . . 11 77 5.3. Existing work . . . . . . . . . . . . . . . . . . . . . . 12 78 6. Research directions . . . . . . . . . . . . . . . . . . . . . 13 79 6.1. Relation to existing research domains . . . . . . . . . . 13 80 6.2. Experimental research objectives . . . . . . . . . . . . . 14 81 7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 14 82 8. Security Considerations . . . . . . . . . . . . . . . . . . . 14 83 9. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . 15 84 10. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . 15 85 11. Informative references . . . . . . . . . . . . . . . . . . . . 15 86 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . 16 88 1. Introduction 90 As currently instantiated, the Internet hour-glass model drives a 91 top-down approach. Current communication networks typically operate 92 with an explicit internal representation of themselves, their network 93 knowledge, and their global goals. Routers follow explicitly 94 (pre-)defined behavior, persistently decide and uniformly execute. 95 Global Internet behavior is evaluated and configuration is when the 96 evaluation indicates that the networking systems are not 97 accomplishing what they were intended to, or when better 98 functionality or performance is expected. 100 In several Internet areas, this operational model shows its limits. 101 Inter-domain routing protocols such as BGP are increasingly impacted 102 by topology and policy dynamics, delaying their convergence due to 103 inherent exploration properties. Network management becomes more and 104 more complex, as networks do not automatically take into account 105 network traffic statistics and other dynamic properties. Several 106 efforts have been undertaken to overcome the increasing number of 107 issues. However, improvement of the routing system to accommodate 108 various scales of challenges in network efficiency, further 109 complicates its operation ([I-D.ietf-idr-bgp-issues]). Further 110 patching the inter-domain routing system and equipment will result 111 into more operational complexity. 113 In this document, we suggest an alternative (bottom-up) approach to 114 the Internet routing and forwarding system operation. Compared to 115 current routed networks requiring explicit specification of expected 116 behavior, self-adaptive systems could dynamically modify or adjust 117 their behavior to varying network conditions in order to tune their 118 operation, optimize their overall performance and even add 119 functionalities through closed-loop adaptive control. 121 We see three main drivers for the development of Learning Capable 122 Communicatino Networks (LCCN): i) the availability of network-related 123 data, ii) the wide range of possible learning paradigms that can be 124 borrowed from domains such as Artificial Intelligence (AI), machine 125 learning, and bio-inspired learning, and iii) the increased CPU 126 capacity available at both forwarding and control plane level, 127 allowing for background monitoring, learning and optimization in 128 routers. 130 The structure of this document is as follows. In Section 2, we 131 describe the opportunities for communication networks to learn how to 132 improve their performance. The next section (Section 3) gives a more 133 formal but broad definition of the concept of learning. Section 4 134 provides a first set of architectural implications of adding learning 135 capability to communication networks. The applicability domain of 136 LCCNs is covered in Section 5, and possible research directions are 137 described in Section 6. Concluding remarks and future work are 138 indicated in Section 9. 140 2. Learning opportunities 142 2.1. Availability of network data and statistics 144 Hosts communicate by sending packets between each other via transit 145 network nodes. As such, a communication network is loaded with 146 packets corresponding to network traffic flows between given network 147 source and destination nodes. Many techniques exist to gather 148 statistics about the resulting traffic flows crossing routers. 150 o Online statistical counters measure properties of transiting 151 traffic in a router using counters, for example the number of 152 packets per destination prefix or used packet size distribution 153 curves 155 o Traffic sampling: instead of counting certain traffic 156 characteristics, unmodified traffic is captured for some time 157 interval. This sample is then used to derive certain 158 characteristics, using e.g. the setting proposed in [Estan04]) by 159 means of sample-and-hold technique. 161 Unfortunately, the resulting statistical data is rarely used to 162 directly improve the routing and/or forwarding decision of network 163 nodes (referring to the active self-adaptive closed control loop in 164 Section 4.1). However, it is clear that network operation could 165 benefit from taking these statistics automatically into account to 166 allow for traffic spreading and network load balancing, ordering of 167 prefix updates in traffic-informed re-routing decisions, and so on. 168 To a lesser extend (since the routing system is deterministically 169 adaptive to topological and/or policy changes), this observation also 170 applies to routing information exchanges. 172 Not only the statistics of network traffic are valuable but also the 173 behavioral aspects of the network itself possibly contain usable 174 information for increasing the performance of the network. 175 Statistics about node or link failures can help network recovery 176 mechanisms to fine tune their operation based on the specific 177 statistical context of the running network. Convergence behavior of 178 routing protocols in the specific running context can be monitored 179 such as to reduce the time of transient loops. In brief, the 180 specific running conditions of communication networks possibly hide 181 (statistical) information, which are currently (largely) unused by 182 current Internet protocols; nevertheless, providing an opportunity to 183 better analyze the behavior of the network behavior depending on the 184 context it is running within. 186 2.2. Availability of processing capacity 188 The possibility of maintaining network statistics is not only 189 dependent on the network conditions and environment themselves, but 190 also on the physical feasibility of monitoring and storing them over 191 longer periods. 193 Supported by Moore's law, we observe that processing power is 194 increasing over last years, either in pure clock frequency of CPU, or 195 in the occurrence of combinations of multiple CPU's on one chip. In 196 combination with the high increase in line card speeds (up to 100 197 Gbps), the possibility of capturing useful network statistics in 198 background seems within reach. 200 3. The learning process 202 Many research fields study the concept of learning from various view 203 points. In the context of LCCNs, learning algorithms correspond to 204 the (broad) class of algorithms that discover the relationship 205 between system variables (i.e. input, output and hidden variables) 206 from data samples of its environment (obtained by means of 207 measurement/monitoring). More formally, the learning process 208 consists of the following steps (see Figure 1). 210 ,--. 211 + + 212 |`--'| 213 |KIB |<------------------+ 214 + + | 215 `--' | 216 | | 217 v | 218 +----------------+ | 219 ,--. | Learner | +------+ 220 E + + | | / \ 221 v |`--'| | +------------+ | / Hypothe- \ 222 e -------->| |------> Learning | |----->\ sis h / 223 n | + + | | algorithm | | \ / 224 t | `--' | +------------+ | +------+ 225 | Training +----------------+ | ^ 226 | data set | | 227 | +--------------------+ | 228 | v + 229 | ,--. +----------------+ / \ 230 | + + | Performer | / \ 231 | |`--'| | +------------+ | / test \ target 232 +-->| |------> Learned | |-------->\ /---> function 233 + + | | hypothesis| | \ / 234 `--' | +------------+ | \ / 235 Test +----------------+ + 236 data set ^ 237 | | 238 +-------------------------------------+ 240 Figure 1 242 o Step 0: Choose training and test data sets associated to a given 243 (sequence of) event(s) observed in the system's running 244 environment 246 o Step 1: Training (learner): learn an hypothesis h (model), 247 function of the input (training data set) that approximates at 248 best output y (symbolic = classification, numeric = regression). 249 Knowledge: use prior "knowledge" stored in Knowledge Information 250 Base (KIB) to learn h 252 o Step 2: Testing (performer): evaluate learned model using test 253 data set 255 4. Architectural implications 257 The control of dynamic systems such as communications networks and 258 routers in particular, can be explained as an interative cycle 259 referred to as the control loop. The coming sections explain the 260 difference of existing communications networks and routers, with the 261 control loop of LCCNs. 263 4.1. From a pre-defined open-loop control towards a self-adaptive 264 closed-loop control 266 The configuration and operation of existing communication networks 267 typically consist of a set of components and algorithms acting in a 268 relatively small space of states, transitions and optimization steps. 269 Let's take as example routers: they distribute topology and/or 270 distance information from which they compute (e.g. shortest) routing 271 paths. Using this information, they derive entries looked up to 272 forward packets based on incoming packets' destination address. When 273 a topological or distance change occurs, routing updates are timely 274 disseminated in the network such that each router achieves a coherent 275 full view of the new network topology and/or distances and can re- 276 compute new routing paths taking into account this new state of the 277 network. While these procedures might seem effective at first sight, 278 they are mostly pre-determined and inflexible with respect to the 279 environment they are running in. 281 Indeed, routers are agnostic to traffic characteristics and to 282 statistics of network failures. This situation occurs because these 283 techniques have been developed in the early days of packet 284 communication networks. At that time, computational and memory 285 resources were scarce, and the resulting techniques needed to act 286 sparingly with the available resources. Moreover, most of these 287 techniques aim to automate manual procedures used to configure or 288 operate communication networks. As such, routers forward packets 289 based on their destination address by applying pre-determined 290 decision rules and execution procedures. 292 While many engineering disciplines, such as the automotive or bio- 293 industry, have adopted learning techniques to improve the performance 294 of their operational control loops, in computer networking, their 295 application has been restricted mainly to passive applications 296 leading to open-loop control procedures. Examples of such 297 applications are: time series models to analyze and predict network 298 traffic data, anomaly detection techniques to check networks for 299 strange events, or statistical models which try to detect Shared Risk 300 Link Groups (SRLG). Most of the applications of learning techniques 301 are used as interesting side information in the context of network 302 operation. They help network managers to understand and predict the 303 behavior of their network; however few existing network operation 304 models include this learning capability into their direct control 305 loop. 307 In this context, the overall objective is to bring the application of 308 data mining and learning techniques one step further: towards the 309 active integration of these techniques into the operational and 310 control processes of communication networks. For instance, we could 311 augment the above control paradigm with a machine learning component 312 enabling the system and network to learn about their own behavior and 313 environment over time, to detect and analyze problems, adapt their 314 decision, and tune their execution using output of models in order to 315 increase their functionality and performance. Systems with such an 316 adaptive closed-loop control have network elements autonomously 317 interrelated and controlled, dynamically adapting to changing 318 environments, and learning desired behavior. These fully distributed 319 and technology-independent systems allow: i) self-configuration and 320 self-organization, ii) self-protection and self-healing, and iii) 321 self-optimization. The objective is to improve the Internet control/ 322 routing and forwarding process by enabling, automating, and 323 distributing the decision making processes involved in their 324 operation. 326 +-----------+ +-----------+ 327 system ==> | analyze |----------->+ decide | <== rules 328 knowledge +-----------+ +-----------+ 329 ^ | 330 | v 331 self- +-----+-----+ +-----------+ 332 monitoring | detect |<-----------+ execute | self- 333 +-----------+ +-----------+ configuration 334 ^ | 335 | v 336 +------------------------------+ 337 | Controlled Element | 338 +------------------------------+ 340 Figure 2 342 Using a more advanced control loop, the routing systems locally learn 343 from network traffic, failure patterns and other context-related data 344 observed in the network, and locally adapt their procedures to 345 optimize their decisions depending on the running context and their 346 internal state. The resulting self-adaptive closed-loop control is a 347 four step cyclic process consisting of: i) a detection phase (e.g., 348 monitor network traffic) which is about monitoring data, ii) an 349 analysis or learning phase (e.g., build traffic models for 350 prediction) in which the data obtained during the detection phase is 351 analyzed and upon which models can be learned, iii) infer rules/ 352 decisions from the performed/learned analysis such that the learned 353 model can influence the operation of the network and iv) an execution 354 phase. 356 4.2. The integration of learning capability 358 While it is premature (and part of the research work) to detail the 359 implications on the Internet architecture, the design of a control 360 system incorporating learning capability would benefit from the 361 following design principles. 363 o Adaptability: modular instead of relying on unified and monolythic 364 approach in order to ensure gradual development (e.g. access vs 365 core router) 367 o Segmentability: rely on relative local view rather than a network 368 global view in order to ensure scalability, robustness, and 369 resiliency 371 o Sizeability: inherits distributed properties and capabilities of 372 routing system (e.g. intra- vs inter-domain) in order to ensure 373 organic deployment --instead of a uniform and ubiquitous plane 374 construction 376 Taking these principles into account, the resulting architecture 377 should specify: i) expected behavior of the self-adaptive closed-loop 378 process, ii) its components, and iii) the interfaces with existing 379 routers' components and between learning-capable routers of a network 380 (both intra- and inter-domain). The resulting closed-loop adaptive 381 control includes a learning component that is either an upfront step 382 or an online process, a feedback phase, and interactions with router/ 383 network control. 385 Today Step 1 Step 2 386 +--------------+ +----------------+ +------------------+ 387 | | | +------------+ | | +--------------+ | 388 | +----------+ | | | Learning | | | | Routing | | 389 | | Routing | | | +------------+ | | | + learning | | 390 | +----------+ | | weak coupling | | +--------------+ | 391 | | ==> | +------------+ | ==> | integrated | 392 | | | | Routing | | | strong coupling | 393 | +----------+ | | +------------+ | | +--------------+ | 394 | |Forwarding| | | +------------+ | | | Forwarding | | 395 | +----------+ | | | Forwarding | | | | + learning | | 396 | | | +------------+ | | +--------------+ | 397 +--------------+ +----------------+ +------------------+ 398 Figure 3 400 Including learning capabilities into current Internet router 401 architectures can follow a phased approach. Internet routers 402 typically consist of two functional components: i) a forwarding 403 component which takes care of processing and forwarding packets 404 according to pre-configured forwarding tables, and ii) a routing 405 component which takes care of distributing topology/distance 406 information, computing (shortest) routing paths using this 407 information, and storing resulting entries into routing tables. 408 Forwarding table entries are subsequently derived from routing table 409 entries. As a first integration step, a new functional component 410 comprising learning capability could be included. The new component 411 would then be weakly coupled to the existing forwarding and routing 412 components. This implies that the routing and/or forwarding 413 component can be enhanced by of the learning component. These 414 functionalities could be called via pre-defined interfaces between 415 the components. While this is an overlaid but modular build-up of a 416 router, integration of learning capability can go one step further. 417 Indeed, in a next phase, instead of a separate learning component, 418 the learning functionality could be tightly integrated into the 419 routing and forwarding components themselves. This implies that the 420 routing and forwarding processes themselves comprise a learning cycle 421 (a self-adaptive closed-loop control). It is clear that both the 422 phasing and the detailed specification of the architecture is an 423 important challenge in the design of LCCNs. 425 4.3. Coexistance with current networking protocols, mechanisms and 426 practices 428 The roll-out of learning capability into communication networks 429 preferrably allows to coexist with well-functioning existing network 430 protocols and mechanisms. This means that LCCNs should not enforce 431 the networking environment to use them or adapt to them, even though 432 they could improve the resulting network performance or solve a 433 number of issues. As such, a transition path towards communication 434 networks including more learning-capability becomes possible without 435 introducing abrupt transition paths. 437 4.4. Complexity/control vs. performance/labour trade-off measurability 439 The implications of using LCCNs should be addressed by determining 440 the relative complexity and understandability they introduce. This 441 does not mean that complex (or black box) LCCN approaches are out of 442 scope, it implies that the additional complexity and 443 understandability resulting from the introduction of this control 444 component should be measurable or can be at least characterized. 445 Measurability (and associated metrics) is an integral part of the 446 investigation work. The assesment should allow users of LCCNs to 447 decide on the level of control vs. performance they are willing to 448 give up/gain. In this context, the analogy can be made with manual 449 configuration of static routing tables vs. running automated shortest 450 path protocols. It is clear that a certain level of control is given 451 up by allowing automated routing protocols to configure routing 452 tables. However, the resulting configuration is verifyable (by 453 routing table inspection), the used algorithm (e.g. Dijkstra 454 shortest path calculation) is known, and the resulting reduction in 455 manual intervention is clear. On the other hand, the more laborous 456 manual configuration allows for setups that are sometimes more tuned 457 to specific traffic patterns (e.g. avoiding bottlenecks) than 458 shortest path-protocols. In most scenario's, the trade-off is clear 459 for network operators: larger networks typically use automated 460 routing protocols for the population of routing tables, whereas 461 smaller, specialised network setups sometimes result into manually 462 configured routing tables. A similar type of trade-off is desired 463 for LCCNs. 465 5. Applicability 467 5.1. Functional domains 469 The incorporation of learning component within the router 470 architecture aims to i) enhance Internet functionality in order to 471 cope with known operational challenges such as manageability, and 472 diagnosability, ii) address new challenges such as security and 473 accountability, and iii) improve its performance (in terms of e.g. 474 scalability and availability) by adapting forwarding and routing 475 system decisions. In this context of network quality, we can think 476 of the automated inclusion of network traffic knowledge into the 477 configuration of routes and resulting forwarding tables. 479 5.2. Scope with respect to the hourglass model 481 Even if learning paradigms can be applied at all levels of the hour- 482 glass model, LCCN-related research focuses on the (largest) lower 483 half of the hourglass model ("everything over IP, and IP over 484 everything"). As depicted in Figure 4, the goal of LCCN research is 485 to apply learning capabilities from the transport layer up to the 486 physical layer (including thus also the network and datalink layers). 488 Whereas learning capability is typically being used at the 489 application layer already, for example by banking applications, 490 large-scale websites such as Amazon or Google, except for TCP, the 491 real networking machinery that is running below is still relying on 492 low-information processes with very limited learning capabilities. 494 The incorporation of a learning component within wired and wireless 495 communication network systems aims to improve both their operation 496 and performance from the physical network layer up to the TCP/IP 497 layer. TCP can be qualified as an exception in the sense that it 498 incorporates some of the procedures involved in learning processes. 499 Indeed, its transmission window size is adaptively changed during the 500 communication between network end points such as to maximize 501 throughput while keeping the resulting congestion as low as possible. 502 However, it mainly concerns end-to-end learning while learning within 503 the network itself provides additional value (as shown by the work 504 performed e.g. in [Tavernier10]). 506 +---------------------+ 507 \ email, WWW, / 508 \ TV, ... / 509 \---------------/ 510 \SMTP,HTTP,RTP/ 511 --- \-----------/ --- 512 ^ \ TCP, / ^ 513 | \ UDP / | 514 | \-----/ | 515 LCCN | / IP \ | 516 scope | /-------\ | 517 | /Ethernet,\ | 518 | / PPP,... \ | 519 | /-------------\ | 520 v / CSMA, Sonet \ v 521 --- /-----------------\ --- 522 /copper,fiber,radio \ 523 +---------------------+ 525 Figure 4 527 5.3. Existing work 529 Although the penetration of learning capability in current network 530 protocols is rather low, in several domains some studies have been 531 conducted on the possible value of introducing learning capability or 532 intelligence into the networking mechanisms. 534 Learning systems have been succesful applied for example in cognitive 535 radio networks and optical networks. Using such systems, wireless 536 network nodes adaptively change their transmission and/or reception 537 parameters to communicate efficiently avoiding interference with 538 other networks and nodes. The adaptive change of these parameters is 539 based on the active monitoring of several factors in the external and 540 internal radio environment, such as radio frequency spectrum, user 541 behavior and network state. More information about cognitive radio 542 networks can be found in [Haykin2007]. 544 [Riziotis07] made a survey on the succesful application of 545 computational intelligence techniques in the domain of photonics and 546 optical networks. Tens of studies are cited on the succesfull 547 application of optimization and learning techniques in the design and 548 operation of optical networks. For example in [Goncalves04], agents 549 make use of Artificial Neural Networks for monitoring an optical link 550 of a network and predicting anomalous situations so that pro-active 551 measures can be taken before faults occur. This technique showed to 552 be significantly less costly compared to providing 1+1 protection on 553 DWDM links. 555 The insight resulting of bringing together conducted research on 556 learning capability in networked environments can result into a 557 common base of and architecture to further investigate and deploy 558 learning capability into new networked contexts. Such a bottom-up 559 approach can be valuable as it can give us lessons in common 560 challenges, and ways to tackle them in order to reach a higher level 561 of adoption of LCCNs. 563 6. Research directions 565 6.1. Relation to existing research domains 567 Learning opportunities in communication networks have characteristics 568 that are typical well-suited for research techniques borrowing from 569 (machine) learning, robotics, AI, computational biology, etc. 571 o Difficult to explicitly characterize: events cannot be well 572 characterized even when examples are available (inherent 573 complexity in characterizing an event) 575 o Correlation: hidden correlations and trends between events within 576 large amounts of associated data 578 o Dynamicity: changing conditions over time (in particular, for 579 routing system but also variability of traffic, user expectations 580 and behaviors) 582 o Quantity: amount of available data is too large for handling by 583 manual intervention 585 o Evolutive: new events are constantly detected/discovered 587 6.2. Experimental research objectives 589 Experimental research is a primary goal of the activities to be 590 conducted. The following objectives would be targeted: 592 o The production of various studies is stimulated and should enable 593 evaluation of performance and functional improvement resulting 594 from the exploitation of various learning paradigms. A common 595 understanding of these paradigms and their associated capabilities 596 could complement this first step. The resulting bottom-up 597 approach allows to combine insights of several use cases involving 598 learning in networks to find the common base and best 599 architecture/practices in the development of LCCNs. 601 o As different distribution models can be considered for what 602 concerns the distribution of the learning processes (taking into 603 account the various objectives but also constraints resulting from 604 network partition), determining which model best fit Internet 605 evolution is a specific target of this research activity. 607 o Iterative cycles of experimentation shall allow to determine 608 suitability of the resulting architecture as well as to determine 609 practical feasibility, applicability and deployability of the 610 concept on a large scale. Documentation of appropriate use cases/ 611 scenarios would complement this work item. 613 7. IANA Considerations 615 This memo includes no request to IANA. 617 8. Security Considerations 619 It is desirable that LCCNs provide visibility on the possible mis-use 620 of their learning capability. As such, the assesment of their 621 attractiveness for deployability becomes easier. 623 Beside the research objectives detailed here above, security 624 mechanisms for "communication channels" between learning components 625 and "learning components" themselves shall be considered comprising 626 among others message authentication but also means to prevent e.g. 627 man-in-the-middle and DDoS attacks. In the LCCN context, the 628 question becomes what is sufficient for protecting the Internet 629 against such attacks. Is it sufficient to provide secure 630 communication channels as well as adequate authentication and 631 verification/validation mechanisms for the information exchanged over 632 these channels, or can we rely on learning to determine protecting 633 decisions systems should take to ensure their own defense against 634 such attacks ? These are security topics that can be further 635 investigated in the context of LCCN research. 637 9. Conclusions 639 Current communication networks fail to use network-related statistics 640 which could be valuable to improve their performance. In addition, 641 current networks fail to provide solutions to challenging issues, 642 because they become too complex to operate and manage by manual/open 643 loop procedures. A learning-capable communication network (LCCN) 644 includes a learning component which learns based on the network 645 environment statistics and adapts and optimizes its behavior upon 646 this. This gives new possibilities to improve network efficiency in 647 several domains including network recoverability, accountability, 648 security, scalability, and so on. The challenge (and next steps) of 649 LCCNs lies into: i) developing self-adaptive closed)loop control 650 system relying on learning capability, ii) building and applying it 651 to various network mechanisms and iii) verifying the resulting 652 prototypes in experimental environments. 654 10. Acknowledgements 656 This work is supported by the European Commission (EC) Seventh 657 Framework Programme (FP7) ECODE project (Grant No.223936). 659 11. Informative references 661 [AI-modern] 662 Russell, S., "Artificial Intelligence: A Modern Approach", 663 2003. 665 [Estan04] Estan, C., "Building a better NetFlow", october 2004. 667 [Goncalves04] 668 Goncalves, C., "Applying artificial neural networks for 669 fault prediction in optical network links", december 2007. 671 [Haykin2007] 672 Haykin, S., "Cognitive radio: brain-empowered wireless 673 communications", february 2007. 675 [I-D.ietf-idr-bgp-issues] 676 Lange, A., "Issues in Revising BGP-4 (RFC1771 to 677 RFC4271)", draft-ietf-idr-bgp-issues-03 (work in 678 progress), August 2010. 680 [PRML] Bishop, C., "Pattern Recognition and Machine Learning", 681 october 2003. 683 [Riziotis07] 684 Riziotis, C., "Computational intelligence in photonics 685 technology and optical networks: A survey and future 686 perspectives", december 2007. 688 [Tavernier10] 689 Tavernier, W., "Using AR(I)MA-GARCH models for improving 690 the IP routing table update", october 2010. 692 Authors' Addresses 694 Wouter Tavernier (editor) 695 Ghent University - IBBT 696 Gaston Crommenlaan 8 bus 201 697 Gent, 9050 698 Belgium 700 Phone: +32(0)9 331 49 81 701 Email: wouter.tavernier@intec.ugent.be 703 Dimitri Papadimitriou 704 Alcatel-Lucent Bell 705 Copernicuslaan 50 706 Antwerpen, 2018 707 Belgium 709 Phone: 710 Email: dimitri.papadimitriou@alcatel-lucent.com 712 Didier Colle 713 Ghent University - IBBT 714 Gaston Crommenlaan 8 bus 201 715 Gent, 9050 716 Belgium 718 Phone: +32(0)9 331 49 70 719 Email: didier.colle@intec.ugent.be