Towards Zero Touch Networks: From the Perspective of Hierarchical Language Systems

In recent years, the proliferation of diversified services, e.g., augmented reality/virtual reality (AR/VR), and emerging network technologies, e.g., 5G/B5G and network digital twins [1], have resulted in a complex and dynamic networking system. Manually managing communication networks is an unacceptable option for network operators. Network autonomy is a clear trend in future that has attracted tremendous attention in the networking community. The “Zero Touch Network” [2] which is to fully integrate the processing of perception, control, adaptivity and assurance to realize the automated management of networks without human intervention, is facing diversified discussions.

Thanks to the great progress made by machine learning, especially deep learning in automating tasks such as protein folding problems [3] and autonomous driving [4], many machine learning methods have also been explored and applied to network automation [5], including resource scheduling [6], anomaly detection [7], proactive defense [8], etc. This provides an opportunity for networks to perform intelligently when facing complex tasks. Despite of the powerful capability of machine learning models on network automation in certain scenarios, they are still designed based on a priori human knowledge and representation of specific tasks with little consideration on the form of intelligence required for zero touch management. This is mainly due to the complexity of network systems which are difficult to generalize the essential properties of network intelligence. Attempting the solution of this challenging problem is an important boost in the development of zero touch networks.

Fundamentally, the premise of realizing a dynamic system that can operate with zero touch is that, there is a mechanism to ensure its self-running behaviors within the bounds of constraints. The behavior trajectories (or phenomena) of the system can be represented explicitly in the form of a language, such as a symbol string marking behavior. Language is a form of external manifestation of the constraints shown inherently in the system. Analogous to the human natural language, the position and semantics of each symbol is constrained by contextual information inherently embedded in human communication. Compared with natural language systems, the form of constraints is more complex and diverse in networking systems. If there is a language system that can be induced by the internal constraints of networking systems, the trajectories (or phenomena) expressed by the language can be seamlessly integrated into the networking system to achieve self-control and management, as shown in the flow depicted by the solid black lines with arrows in Fig. 1.

With the ultimate goal of realizing zero touch networks to autonomously manage the network, three key constraints need to be considered in the process of developing reasonable solutions. Below shows the details of the three constraints in networking systems:

• •

  Spatial constraint: A network is composed of a series of entities.

• •

  Temporal constraint: Network tasks are sequential operations performed on network entities.

• •

  Causal constraint: Performing an action causes state changes of the network.

To incorporate these constraints in the design of language systems for zero touch networks, there needs to be a way to construct these constraints as grammar rules of the network language, so that the language derived from the grammar rules can be naturally adaptable and executable in the zero touch network. These three constraints ensure that a language conforms to the semantics of the network domain.

In this article, we explore the zero touch network from the perspective of language-like systems that 1) reveal the three constraints in the form of a language system, and use Spatial-Temporal-Causal AND-OR Graph (STC-AOG) as the grammatical structure of the language system; 2) given a network task, the Spatial-Temporal-Causal Parsing Graph (STC-PG) can be obtained by parsing the STC-AOG, so as to know what the network task is, how to perform it, and what network intents will be affected; 3) provide network operators with an intuitive understanding of the current network using the translation of STC-AOG or STC-PG into first-order languages.

The overall goal of zero touch networks is to enable networks to become more autonomous and allow network operators to confidently entrust them with complex, tedious tasks. This fascinating vision has attracted continued attention from the research community, standards organizations and network providers. European Telecommunications Standards Institute (ETSI) established the Zero touch network and Service Management Industry Specification Group (ZSM ISG). Its goal is to eventually automate 100% of network operations processes. The group has published 10 standards to define an end-to-end operational framework for the future of intelligent networks, enabling networks to be agile, efficient and automated [9]. Huawei and Juniper have each released their zero touch network related products, Autonomous Driving Networks (ADN) [10] and Self-Driving Networks [11], respectively. The ADN has produced a three-layered open architecture delivering intelligence for networks, which are cloud intelligence for data training as well as model generation and optimization, network intelligence for service intent automation, and network element intelligence for providing short period awareness analysis and inference capabilities. The self-driving network aims to ease the heavy lifting of network administrators by self-configure, monitor, manage, correct, defend and analyze with little human intervention.

Artificial Intelligence (AI) powered by machine learning and deep learning technologies has been considered as a key enabler for zero touch networks. Alongside standards organisations and network providers, research community is progressing AI initiatives looking at autonomous networks. Chergui et al. [2] proposed a framework with a distributed and AI-driven management and orchestration system for large-scale deployment of network slicing in the sixth-generation network (6G). From the perspective of automata learning, Luque-Schempp et al. [12] built an intelligent controller based on learning from the live network, which solves the problem of dynamically predicting and applying the appropriate 5G non-public networks configuration to meet the requirements of the time sensitive network traffic. The authors in [13] introduced a zero touch 6G large-scale network slice analysis engine based on statistical federated learning, which performs slice-level resource prediction through offline learning. Shaghaghi et al. [14] proposed a zero touch, deep reinforcement learning based proactive failure recovery framework for stateful network functions virtualization enabled networks.

The above work has proposed solutions in various scenarios, but lacks a unified intelligent expression model to express the intent of network tasks, what the task is to do and how to do it. This article manages to make up for this lack from the perspective of language-like systems.

In this section, we discuss three important issues of the hierarchical language framework, as shown by the three light green boxes in Fig. 1. The first (bottom) one is the language system. This part mainly discusses the grammar, symbols and semantics, which are the building blocks of the language system. The second (middle) one is to discuss how to use the language system to construct a unified framework for understanding networking systems. The last (top) one discusses the logical description produced by the network language system to provide network engineers/operators with an understandable description of the current network. It is an interface between the network language system and the human language system.

To verify the capability and feasibility of the proposed the hierarchical language framework, we present a networking scenario as an example: task offloading in edge computing. Task offloading is a common application in the field of Internet of Things and edge computing. When the user equipment cannot provide sufficient computing resources, the task is offloaded to a suitable edge server or a cloud center for auxiliary service.

The considered scenario is shown in the bottom part of Fig. 4. The road-side unit (RSU) and edge server (ES) are deployed at the road side. ES provides nearby computing services for vehicular applications, and RSU assists vehicular applications in terms of forwarding tasks to the cloud through the WAN. Considering the movement of vehicles, all RSUs and ESs are connected through the MAN to achieve roaming. In addition, vehicular applications can also communicate with roadside base stations and offload tasks to the cloud via the cellular mobile network. Therefore, a task has three options for offloading: the first one is to offload to ES; the second one is to offload to the cloud through RSU; and the last one is to offload to the cloud through cellular mobile networks. We deploy probes on the ports of network entities to collect path data for task offloading. The path data is shown as the network behavior trajectory in the blue box in Fig. 4. Each record represents a path. Paths are represented by strings of probe symbols, which denote the flow of traffic through these entities. According to the path data, all possible combinations of the two probe symbols are formed to construct a set of binary grammar fragments, as shown on the left side of the blue box in Fig. 4. These fragments are the basic blocks to build grammatical structures of the S-AOG and T-AOG. The root node of the fragment represents a possible micro-action, and the child nodes denote the entities involved in the micro-action. Each record is treated as a complete operation that contains multiple micro-actions.

The learning of STC-AOG is to obtain a compact grammar structure that can explain the path data. In the experiment, the normalized exponential function is used as the prior probability of the grammar structure, which is proportional to the negative of the number of nodes of the grammar structure. Hence, it serves as an indicator of compactness. The more compact the grammar structure, the larger the prior probability. In order to avoid blind pursuit of compactness, which reduces the ability of grammar to explain data, likelihood probability is used to describe the probability of data under a given grammar structure. As shown in the lower figure of the experimental results box (the pink box) in Fig. 4, as the number of inductive iterations increases, the prior and likelihood rise first and then stabilize, indicating that the algorithm can find an optimal compact structure that conforms to the path data. The grammar structure is shown in the T-AOG and S-AOG boxes in the figure. From the experimental results, it can be seen that the number of fragments of S-AOG is reduced to 7 (compared to 28 fragments of the binary grammar fragments). The $A_{1}$, $A_{5}$, $A_{7}$ segments in S-AOG have OR nodes, which increase the compactness of S-AOG. The learned T-AOG has three operations, $O_{c}$, $O_{r}$ and $O_{e}$, which respectively indicate that computing tasks are offloaded to the cloud through the cellular network, offloaded to the cloud through the RSU, and directly offloaded to the ES. These three operations cause the $F_{1}$ feature to change, where $V_{c}$ means the task is executed in the cloud, and $V_{e}$ represents the task is executed at the edge.

In the experiment, we collect 18 features and the corresponding meaning of these features are shown in Table I, of which the feature $F_{0}$ is the intent feature. It indicates that network operators want to maximize the success rate of task offloading (task execution completes and successfully returns to the user device). The upper figure of the experimental results box (the pink box) in Fig. 4 is the visualization of the 18 features in the semantic space. The color bar on the right is the color represented by feature $F_{0}$ to feature $F_{17}$, respectively. It can be seen that the color of the feature $F_{0}$ is fused with the colors of the feature $F_{1}$ and the feature $F_{2}$. This reveals that the semantics of these three features are highly correlated so that $F_{0}$ is the root node, and $F_{1}$ and $F_{2}$ are its child nodes in C-AOG. Based on the features semantic learning for C-AOG and the prior and likelihood probabilities learning for S-AOG and T-AOG, the experiment yields a grammar structure STC-AOG of computational offloading tasks for edge computing scenarios.

For each path data, STC-PG can be parsed from the learned STC-AOG, and then translated into first-order logic. For example, the record {USER, CC, CLOUD, CC, USER} can be interpreted into the first-order logic as follow:

1. 1.

   $F_{1}(V_{c})\wedge F_{2}(*)\to Intent$

2. 2.

   ${\forall}O_{c}~{}(V_{c}(O_{c}))$

3. 3.

   $A_{1}(CC,CLOUD)\wedge A_{5}(CC,USER)\to O_{c}$

This first-order logic statement visualizes the relationship between the intent (what to do) and a specific task (how to do). The first sentence shows that the value of $F_{1}$ changing to $V_{c}$ will affect the realization of the intent. The second sentence further indicates that only this operation is the reason for changing the state of feature $F_{1}$ to $V_{c}$. The last sentence explains in detail how the operation is performed. The micro-action $A_{1}$ is executed on $CC$ and $CLOUD$, and the micro-action $A_{5}$ is executed on $CC$ and $USER$. These two micro-actions are executed in sequence to complete $O_{c}$ operations. It means the task is offloaded to the cloud and backtracked to the user device.

To verify the capability of the proposed language system in network optimization, we conduct additional experiments. According to the previous experiments, the intent feature $F_{0}$ is related to the feature $F_{1}$. Three operations $O_{c}$, $O_{r}$ and $O_{e}$ can affect the value of the $F_{1}$. We utilize the previous semantic learning model as shown in Fig. 3 to learn the semantic relationship between the intent feature $F_{0}$ and the entities involved in $O_{c}$, $O_{r}$ and $O_{e}$ operations (an entity is represented by its attributes). We obtain the entity attributes that are most relevant to the intent feature. Based on this, we design a decision model, where the input is an entity attribute that is semantically strongly related to the intent feature, and the output is the probability distribution of the three operations. The operation with the highest probability indicates that the model predicts that it is the most likely operation to achieve the optimization intent.

The experimental results are shown in Fig. 5, and the purple line is the performance of the stochastic decision model on the intent of the task failure rate. The decision models of the red and green lines use the data generated by the stochastic decision model as training data. The input of the decision model of the red line is the features designed based on human prior knowledge, and that of the decision model of the green line is the feature that is strongly related to the intent learned by our proposed language system. Experimental results show that our decision model performs optimally. It can correct the bias introduced by the artificially designed model, so that the input features of the model are more in line with the network intent.

Many studies have been reported for zero touch networks. In this article, our solution is proposed by building a language system which is specifically designed for the networking domain. In this section, we discuss research challenges and open issues which can guide the further research on this topic.

This article proposed a hierarchical language framework for zero touch networks. In the language system of the framework, the spatial, temporal and causal constraints endogenously in the network system are expressed explicitly by the grammar structure of STC-AOG. This article discussed the initializing and learning the STC-AOG, the parsing of STC-AOG to obtain STC-PG, and the transformation of hierarchical STC-AOG into a flattened symbol string in the form of first-order logic. Given a task, the zero touch network knows what the network task is, how to perform it, and what intents will be affected with the help of the network language system. In addition, symbols strings induced from the network language system provides network operators with an intuitive understanding of the network. A case study was developed to demonstrate the feasibility of the proposed hierarchical language framework for zero touch networks by performing the scenario of task offloading in edge computing. At last, we provided a list of research challenges and open issues that can be useful to the community for carrying out further research.