Learning Capable Communication Networks (LCCN)
o) Scope and Coverage
Control processes of current communication networks are determined at design time following their global objective of minimizing the cost/performance ratio for a given set of pre-determined functionality. These networks are configured so as to meet pre-defined goals known to be met by pre-determined configuration parameter setting. Once configured, communication systems such as routers follow explicitly pre-defined behavior, persistently decide and uniformly execute. The off-line evaluation of their global behavior leads to non-automated changes of their configuration (re-configuration) when evaluation results indicate that the communication systems are not accomplishing what they were intended to do, or when better functionality or performance is possible. The consequences are that unexpected events (due to external/environmental or internal state) change lead to sub-optimal behavior that can only be fixed by patching or sometimes by reconfiguration.
The control of dynamic systems such as communications networks, and routers in particular, can be explained as an interactive cycle referred to as the control loop. While open loop and non-automated control has been effectively used since so far, this paradigm is progressively showing its limits. On one hand, the underlying procedures are mostly pre-determined and inflexible with respect to the variation of their environmental properties; on the other hand, as the network scales, they become also more complex and thus more expensive to operate. Using a more advanced and distributed control loop, individual routers for instance can learn e.g. from network traffic and failure patterns as well as other context-related data observed in the network, and adapt their procedures to optimize their decisions depending on the running context and their internal state.
The overall objective of this group is to bring the application of data mining and learning techniques towards their integration into the control processes of communication networks, and in particular to i) Routing (incl. traffic engineering, resiliency, stability, etc.), ii) Network self defense (incl. distributed intrusion/attack detection, distributed anomaly detection, etc.), iii) Control of optical networks, and iv) Mobile/ad-hoc/sensor networks. Realization of this objective involves the introduction of a distributed closed-loop control process which automatically adapts its decisions and its executions by monitoring/diagnosing its internal state, activity/behavior, and environment over time. Automated adaptation can be enabled by applying various learning techniques. Learning provides the means to i) detect, identify, and analyze problems, ii) determine how and when (timely) to adapt decisions and executions and iii) decide when to operate autonomously or in cooperation. The resulting self-adaptive closed-loop control includes a learning component that is either an upfront step or an online process, a feedback phase, and interactions with router/network control.
The addition of a learning component as part of the control system of wired and wireless communication networks aims to improve their functionality and performance from the physical network layer up to the TCP/IP layer. Indeed, this component is expected to i) enhance network controlability and diagnosability, ii) address important challenges such as security (self-defense) and accountability, and iii) improve network performance (e.g. resource usage) and resiliency by dynamically adapting forwarding and routing system decisions.
o) Research Activities
To reach this objective, iterative cycles of experimental research will be stimulated and conducted on representative use cases allowing to evaluate various learning techniques as well as to determine their developability and executability on a large scale basis.
The initial set of work items comprises (for each item, goal(s) are identified outlining the expected results and outputs):
Goal: achieve common understanding of the various definitions and terms that will be used throughout this research phase.
Goal: determine the value and benefits (as well as possible trade-offs) when applying learning techniques to well-identified challenges for the above-mentioned domains; document these techniques, their running conditions and their applicability.
Goal: identify the main areas of potential evolution of control principle(s) and model(s) compared to current control design; document the architectural implications and potential issues that may arise when applying the experimented learning techniques.
Goal: determine possible generalization and cross-domain applicability of the investigated learning techniques for communication networks so as to layout the specification of a distributed learning component.
Note: the scope of the work does not include management plane related research activities
Expected duration for the realization of these objectives is about 30 months. Depending on the results obtained from this initial set of work items, subsequent research work may include the following items:
o) Proposed Organization
This group aims to provide a forum for exchange of ideas and results among researchers interested in distributed control as well as in customizing learning paradigms and techniques to network control.
Most of the communication inside the group will be organized by means of mailing lists. The group will hold regular physical meetings (at least once a year) in conjunction with IETF meetings. Additional meetings may be held at IETF or other venues, such as in conjunction with related conferences and workshops.
The group will produce Informational and Experimental RFCs in order to document its activity and to formalize the outcome of the research topics carried. Such documentation could become input to IETF working groups. The group will also encourage experimentation of learning techniques to validate this exploratory research.
o) Relationship to other IRTF Research Groups and IETF Working Groups
The group operates in an open fashion (meetings and mailing list).