Cognitive architectures emerge from and for the network, seeking to apply AI/ML to novel closed control loops for network management [Boutaba et al., 2018] to tackle the operational adversities by nature of networked systems, meaning constant changes in KPIs as much as in infrastructure conditions – i.e., the impossibility of exact solutions feasible for management. Lately, virtualization models (SDN/NFV) and increased connectivity amplified the complexity of such nature, turning networks critically hard to tame and feasibly prone to automation. Root causes of those issues include large telemetry datasets, a myriad of heterogeneous control interfaces, complex protocol behaviors, and increased specificity of user traffic profiles (a.k.a. slicing). Among the many fold research topics involved in cognitive architectures and AI-powered mechanisms [Letaief et al., 2019], some aspects pose more relevance to the scope of SMARTNESS: (i) Autonomy – establishes policy-based decision making to realize intents within reliable operational margins; (ii) Dynamics – comprises stable operations to support cognitive workflows (e.g., observe, orient, decide, act), given environment adversities and variable key performance indicators (e.g., stability, goodness, resiliency) while in production; and (iii) Federation – stands for composing an associative stratum among cognitive architectures, its components, data, and applications, contributing with each other in different environments, scales and workflows. Approaching the enlisted topics, SMARTNESS will delve into the following STA topics to realize cognitive architectures for communication networks and their services:
- Scalable models: to establish composable building blocks for algorithms, control-loops, distributed learning and cognitive architectures as a whole. Ideally cognitive models need to detain the ability to be implemented as a system of systems, enabling federation for its components, via flexible interfaces;
- Intent resolution: aims to conduct research on flows of information top-down and bottom-up, specifically on how a intent definition can be decomposed into the configuration of algorithms (e.g., hyperparameters) and programmable behaviors (e.g., features selection) to attain predictable operational targets;
- Policy-oriented decision making: to tackle the formal definition of policies and their specification through adequate languages, semantics and behaviors (i.e., its modifiable and interactive aspects). While governed by policies, cognitive architectures need to resolve problems within reliable operational margins, taking decision focused on maintaining stationary states;
- Survivability: aims to enable resilient continuous operation based on self-healing mechanisms for closed control loop automation, specially targeting fault tolerance and the stability of critical service KPIs in real-time. The core of this target aims to focus on the autonomy of cognitive architectures;
- Adaptability: consists in developing cognitive models applicable to different scenarios via transfer learning, and adjustments on-the-fly of service capabilities according to the demands of cognitive architecture components. Those concern features highly needed in constantly evolving and dynamic cognitive architectures;
- Data provenance: cope with continuous adaptability of algorithms for evolving networking conditions by keeping them fed with the proper sources of data to maintain operational KPIs (e.g., accuracy). While in operation, a cognitive architecture needs mechanisms to self-tune its behavior, adapting to changes in features according to the sources of data available;
- Heterogeneous and cross-domain interfaces: cognitive architectures depend on adaptable interfaces to attain programmability and modifications on its behavior dynamically. Within multiple technological domains such capabilities must be satisfied, possibly explored and developed via semantic models. The design and construction of open communication models and APIs should support multiple administrative domains operations to reach, share, trust and execute cognitive architecture components, their sources of data, the data itself, algorithms, decisions, actuation models, etc;
- Network digital twins for advanced DevOps: Cloud-network development and operations (DevOps) could leverage network digital twins as virtual models of processes, products, people, places, systems, devices, services, or states of any real-world entity linked to the real object via a digital thread. A network digital twin could capture both historical, present, and predictive information and associations of the modelled entities, to drive inferences and insightful analytics. This pairing of the virtual and physical worlds through the digital thread allows analysis of monitored data from systems to anticipate and manage problems before they even occur; prognoses systems to prevent downtime; optimizes and develops new business opportunities; and plans future business activities by using simulations and AI-based analytics [Fuller et al., 2020].