FIG. 3 illustrates an example network assurance system 300 (e.g., through execution of network assurance process 248 by one or more devices in a network), according to various embodiments. As shown, at the core of network assurance system 300 may be a cloud-based network assurance service 302 that leverages machine learning in support of cognitive analytics for the network, predictive analytics (e.g., models used to predict user experience, etc.), troubleshooting with root cause analysis, and/or trending analysis for capacity planning. Generally, network assurance system 300 may support both wireless and wired network, as well as LLNs/IoT networks.

In various embodiments, cloud service 302 may oversee the operations of the network of an entity (e.g., a company, school, etc.) that includes any number of local networks. For example, cloud service 302 may oversee the operations of the local networks of any number of branch offices (e.g., branch office 306) and/or campuses (e.g., campus 308) that may be associated with the entity. Data collection from the various local networks/locations may be performed by a network data collection platform 304 that communicates with both cloud service 302 and the monitored network of the entity.

The network of branch office 306 may include any number of wireless access points 320 (e.g., a first access point AP1 through n^thaccess point, APn) through which endpoint nodes may connect. Access points 320 may, in turn, be in communication with any number of wireless LAN controllers (WLCs) 326 (e.g., supervisory devices that provide control over APs) located in a centralized datacenter 324. For example, access points 320 may communicate with WLCs 326 via a VPN 322 and network data collection platform 304 may, in turn, communicate with the devices in datacenter 324 to retrieve the corresponding network feature data from access points 320, WLCs 326, etc. In such a centralized model, access points 320 may be flexible access points and WLCs 326 may be N+1 high availability (HA) WLCs, by way of example.

Conversely, the local network of campus 308 may instead use any number of access points 328 (e.g., a first access point AP1 through n^thaccess point APm) that provide connectivity to endpoint nodes, in a decentralized manner. Notably, instead of maintaining a centralized datacenter, access points 328 may instead be connected to distributed WLCs 330 and switches/routers 332. For example, WLCs 330 may be 1:1 HA WLCs and access points 328 may be local mode access points, in some implementations.

To support the operations of the network, there may be any number of network services and control plane functions 310. For example, functions 310 may include routing topology and network metric collection functions such as, but not limited to, routing protocol exchanges, path computations, monitoring services (e.g., NetFlow or IPFIX exporters), etc. Further examples of functions 310 may include authentication functions, such as by an Identity Services Engine (ISE) or the like, mobility functions such as by a Connected Mobile Experiences (CMX) function or the like, management functions, and/or automation and control functions such as by an APIC-Enterprise Manager (APIC-EM).

During operation, network data collection platform 304 may receive a variety of data feeds that convey collected data 334 from the devices of branch office 306 and campus 308, as well as from network services and network control plane functions 310. Example data feeds may comprise, but are not limited to, management information bases (MIBS) with Simple Network Management Protocol (SNMP) v2, JavaScript Object Notation (JSON) Files (e.g., WSA wireless, etc.), NetFlow/IPFIX records, logs reporting in order to collect rich datasets related to network control planes (e.g., Wi-Fi roaming, join and authentication, routing, QoS, PHY/MAC counters, links/node failures), traffic characteristics, and other such telemetry data regarding the monitored network. As would be appreciated, network data collection platform 304 may receive collected data 334 on a push and/or pull basis, as desired. Network data collection platform 304 may prepare and store the collected data 334 for processing by cloud service 302. In some cases, network data collection platform may also anonymize collected data 334 before providing the anonymized data 336 to cloud service 302.

In some cases, cloud service 302 may include a data mapper and normalizer 314 that receives the collected and/or anonymized data 336 from network data collection platform 304. In turn, data mapper and normalizer 314 may map and normalize the received data into a unified data model for further processing by cloud service 302. For example, data mapper and normalizer 314 may extract certain data features from data 336 for input and analysis by cloud service 302.

In various embodiments, cloud service 302 may include a machine learning (ML)-based analyzer 312 configured to analyze the mapped and normalized data from data mapper and normalizer 314. Generally, analyzer 312 may comprise a power machine learning-based engine that is able to understand the dynamics of the monitored network, as well as to predict behaviors and user experiences, thereby allowing cloud service 302 to identify and remediate potential network issues before they happen.

Machine learning-based analyzer 312 may include any number of machine learning models to perform the techniques herein, such as for cognitive analytics, predictive analysis, and/or trending analytics as follows:

- Cognitive Analytics Model(s): The aim of cognitive analytics is to find behavioral patterns in complex and unstructured datasets. For the sake of illustration, analyzer 312 may be able to extract patterns of Wi-Fi roaming in the network and roaming behaviors (e.g., the “stickiness” of clients to APs 320, 328, “ping-pong” clients, the number of visited APs 320, 328, roaming triggers, etc.). Analyzer 312 may characterize such patterns by the nature of the device (e.g., device type, OS) according to the place in the network, time of day, routing topology, type of AP/WLC, etc., and potentially correlated with other network metrics (e.g., application, QoS, etc.). In another example, the cognitive analytics model(s) may be configured to extract AP/WLC related patterns such as the number of clients, traffic throughput as a function of time, number of roaming processed, or the like, or even end-device related patterns (e.g., roaming patterns of iPhones, IoT Healthcare devices, etc.).
- Predictive Analytics Model(s): These model(s) may be configured to predict user experiences, which is a significant paradigm shift from reactive approaches to network health. For example, in a Wi-Fi network, analyzer 312 may be configured to build predictive models for the joining/roaming time by taking into account a large plurality of parameters/observations (e.g., RF variables, time of day, number of clients, traffic load, DHCP/DNS/Radius time, AP/WLC loads, etc.). From this, analyzer 312 can detect potential network issues before they happen. Furthermore, should abnormal joining time be predicted by analyzer 312, cloud service 302 will be able to identify the major root cause of this predicted condition, thus allowing cloud service 302 to remedy the situation before it occurs. The predictive analytics model(s) of analyzer 312 may also be able to predict other metrics such as the expected throughput for a client using a specific application. In yet another example, the predictive analytics model(s) may predict the user experience for voice/video quality using network variables (e.g., a predicted user rating of 1-5 stars for a given session, etc.), as function of the network state. As would be appreciated, this approach may be far superior to traditional approaches that rely on a mean opinion score (MOS). In contrast, cloud service 302 may use the predicted user experiences from analyzer 312 to provide information to a network administrator or architect in real-time and enable closed loop control over the network by cloud service 302, accordingly. For example, cloud service 302 may signal to a particular type of endpoint node in branch office 306 or campus 308 (e.g., an iPhone, an IoT healthcare device, etc.) that better QoS will be achieved if the device switches to a different AP 320 or 328.
- Trending Analytics Model(s): The trending analytics model(s) may include multivariate models that can predict future states of the network, thus separating noise from actual network trends. Such predictions can be used, for example, for purposes of capacity planning and other “what-if” scenarios.

Machine learning-based analyzer 312 may be specifically tailored for use cases in which machine learning is the only viable approach due to the high dimensionality of the dataset and patterns cannot otherwise be understood and learned. For example, finding a pattern so as to predict the actual user experience of a video call, while taking into account the nature of the application, video CODEC parameters, the states of the network (e.g., data rate, RF, etc.), the current observed load on the network, destination being reached, etc., is simply impossible using predefined rules in a rule-based system.

Unfortunately, there is no one-size-fits-all machine learning methodology that is capable of solving all, or even most, use cases. In the field of machine learning, this is referred to as the “No Free Lunch” theorem. Accordingly, analyzer 312 may rely on a set of machine learning processes that work in conjunction with one another and, when assembled, operate as a multi-layered kernel. This allows network assurance system 300 to operate in real-time and constantly learn and adapt to new network conditions and traffic characteristics. In other words, not only can system 300 compute complex patterns in highly dimensional spaces for prediction or behavioral analysis, but system 300 may constantly evolve according to the captured data/observations from the network.

Cloud service 302 may also include output and visualization interface 318 configured to provide sensory data to a network administrator or other user via one or more user interface devices (e.g., an electronic display, a keypad, a speaker, etc.). For example, interface 318 may present data indicative of the state of the monitored network, current or predicted issues in the network (e.g., the violation of a defined rule, etc.), insights or suggestions regarding a given condition or issue in the network, etc. Cloud service 302 may also receive input parameters from the user via interface 318 that control the operation of system 300 and/or the monitored network itself. For example, interface 318 may receive an instruction or other indication to adjust/retrain one of the models of analyzer 312 from interface 318 (e.g., the user deems an alert/rule violation as a false positive).

In various embodiments, cloud service 302 may further include an automation and feedback controller 316 that provides closed-loop control instructions 338 back to the various devices in the monitored network. For example, based on the predictions by analyzer 312, the evaluation of any predefined health status rules by cloud service 302, and/or input from an administrator or other user, controller 316 may instruct an endpoint client device, networking device in branch office 306 or campus 308, or a network service or control plane function 310, to adjust its operations (e.g., by signaling an endpoint to use a particular AP 320 or 328, etc.).

As noted above, in a traditional network support model, the onus is largely placed on users to identify and report networking issues via a ticketing system. More specifically, when a user encounters a problem (e.g., a lack of connectivity, degraded performance, etc.), the user is then expected to submit a support ticket that is then routed to support personnel for review. Because of the delay in this process, some systems have employed the use of chatbots to walk a user through a series of steps or questions, to help resolve their issue. However, such chatbots are largely based on Natural Language Processing (NLP) and pattern matching, making them very limited in terms of their ability to handle complex queries and queries for which they haven't been explicitly programmed to respond.

Indeed, the NLP capabilities of chatbots and virtual assistants for network assurance today are largely limited to making unstructured search queries possible. Consider, for instance, the following query:

Troubleshoot Client abc-123

Here, the NLP merely consists in allowing the user to use a synonym of “troubleshoot” and matching the right token to some pre-defined filters or commands. For instance, the above query may trigger a search for all issues affecting the client with hostname abc-123. The following would have led to the same outcome:

List Problems for Client abc-123

In practice, the most advanced “AI” capabilities offered by chatbot services such as Tidio are even more limited, as the bot operator must pre-define a list of “likely sentences” that the visitor might say to trigger a pre-defined message.

Recently, however, conversational models based on large language models (LLMs) such as ChatGPT by OpenAI, LaMDA by Google Brain, LLAMA by Meta AI, and Sparrow by DeepMind have significantly raised the bar with respect to users' expectations when interacting with chatbots and virtual assistants. As would be appreciated, LLMs can not only perform complex matches from a semantic perspective, but also generate complex sentences that are not pre-scripted. However, implementing such a conversational model is not straightforward, as networks are highly dynamic, complex systems.

Conversational Network Assurance Using Large Language Models

The techniques herein introduce an architecture that allows for the use of Large Language Models (LLMs) in conjunction with a network assurance service, to aid end users, as well as network support personnel, to identify and resolve network issues. In some aspects, a conversational model is also introduced herein that serves as a gateway between these two sets of users and is able to generate output text that goes beyond simply pre-programmed sentences in a scripted scenario.

Specifically, according to one or more embodiments, a device receives, at a first large language model executed by a device, textual input from a user of a network regarding a networking issue in the network. The device issues, by the first large language model and to a second large language model, one or more questions regarding the network based on the textual input. The device receives, at the first large language model and from the second large language model, one or more answers to the one or more questions. The device generates, by the first large language model, a textual response to the textual input for presentation to the user.

Illustratively, the techniques described herein may be performed by hardware, software, and/or firmware, such as in accordance with the network assurance process 248, which may include computer executable instructions executed by the processor 220 (or independent processor of interfaces 210) to perform functions relating to the techniques described herein.

Operationally, FIG. 4 illustrates an example architecture 400 for conversational network assurance using large language models, according to various embodiments. At the core of architecture 400 is conversational assurance process 249, which may be executed by a controller for a network, a networking device, a service, or another device in communication with a network to be monitored. For instance, conversational assurance process 249 may be executed by a controller for a network, a particular networking device in the network (e.g., a router, a firewall, etc.), another device or service in communication therewith, or the like, to perform the techniques herein. In addition, conversational assurance process 249 may be executed in conjunction with network assurance process 248, in some embodiments, as described below.

As shown, application experience optimization process 248 may include any or all of the following components: a conversational agent 402 and a network question and answer agent 404. As would be appreciated, the functionalities of these components may be combined or omitted, as desired. In addition, these components may be implemented on a singular device or in a distributed manner, in which case the combination of executing devices can be viewed as their own singular device for purposes of executing conversational assurance process 249.

According to various embodiments, conversational agent 402 may take the form of a conversational agent, referred to herein as a “Network Assurance AI (NAAI),” based on a suitable language model, such as an LLM. In various embodiments, the NAAI may mediate the chat-based interactions between 1.) end users in the network, 2.) the network itself, and 3.) the network operators/network support personnel. To this end, training of conversational agent 402 can be achieved as follows:

It mediates all exchanges between the users, the network team, and the network itself. To this end, we introduce two key novelties:

- 1. Train the model of conversational agent 402 to mediate and enrich a conversation between two parties, each composed of possibly multiple people, as opposed to having a dialog with only one person. In this case, the NAAI converses both with one or more users (inquiring about the symptoms they face and the criticality of the problem) and members of the network team (providing insights into the user issue augmented with relevant facts from the network).
- 2. Instead of limiting the model input to a mere sequence of tokens (i.e., natural language), we also condition its input on the state of the network, and we allow it to issue so-called queries to a second model/agent implemented by network question and answer agent 404, referred to herein as the “Network Question Answerer (NQA),” whose role is to act as form of attention mechanism, as detailed below.

In a primary embodiment, the NAAI implemented by conversational agent 402 may be based on a LLM (e.g., GPT-3, GPT-3.5, etc.) and may be re-trained in multiple phases, such as the following:

- 1. Train a supervised policy using demonstration data: such data could come from former technical support cases, transcripts from support conversations, and/or one or more existing knowledge bases (e.g., Cisco DNA Center's Machine Reasoning Engine base, existing bot scripting information from the Cisco Networking Bot, the Cisco CX bot, etc.). Here, prompt engineering could be leveraged with model retraining to make the NAAI aware of its persona, i.e., 1.) a first persona used to interact with an end user as a tech support bot or 2.) a second persona used to interact with the network team (e.g., network support personnel) as an assurance service. To this end, all transcripts may be labeled with information distinguishing between the two, such as with a prefix that indicates “User speaking” or “Network engineer speaking” and used in the transcription.
- 2. Train two reward models, one for each of the two personas above, to rank outputs from best to worse: here, the system may ask both users and network experts (either in live settings or using crowdsourcing) to rank or label sample outputs.
- 3. Optimize a policy against the reward model obtained in step 2: this step requires a reinforcement learning such as Proximal Policy Optimization.

This initial training phase results in a model that can mediate the conversation between a user and a network team, such as the following interactions:

TABLE 1

User-NAAI
Network engineer-NAAI

User A> Hello, I am facing some

issues with Webex.

NAAI> Hello, thanks for reporting

your issue.

NAAI> Hello, User A faces some

issues with Webex.

Engineer 1> Yes, he is in building

X and we currently have an ISP

issue. Tell him that this is going to

be resolved in 5 minutes.

NAAI> Your issue is going to be

resolved in 5 minutes. The Internet

Service Provider in your area is

currently experiencing a disruption.

We are sorry for the inconvenience.

User A> Thank you.

NAAI> You're most welcome.

Furthermore, if a second user B asks the same question, the bot could handle it on its own:

TABLE 2

User-NAAI
Network engineer-NAAI

User B> Hello, Webex doesn't work for

me.

NAAI> Hello, thanks for reporting your

issue. Are you located in building X?

User B> Yes.

NAAI> Other users have reported the

same issue, and it should be resolved

within a few minutes. The Internet

Service Provider in your area is currently

experiencing a disruption. We are sorry

for the inconvenience.

Of course, the notable shortcoming in the above conversation is that the NAAI must ask whether User B is located in building X, which is a piece of information that is easy to obtain from the network itself. Similarly, the following interaction is not possible without access to further information:

TABLE 3

User-NAAI
Network engineer-NAAI

Engineer 1> Yes, he is in building X and

we currently have an ISP issue. Tell him

and all other active Webex users in this

building that this is going to be resolved

in 5 minutes.

If the bot is to execute such an instruction, it must 1.) list all of the users in the building, and 2.) look up users that have active meetings. This is where the Network Question Answerer (NQA) comes into play.

According to various embodiments, network question and answer agent 404 may take the form of a second LLM or other suitable language model that implements a second agent, the NQA. Like the NAAI, the NQA may also be trained to interact under two different personas:

- 1. In its first persona, it takes a (natural language) question as input and translate it into a query to be executed against a data warehouse (e.g., a database, datalake, etc.) with a well-defined schema (e.g., using a distributed SQL query engine such as Trino or Presto).
- 2. In its second persona, it produces an answer based on the query results.

To do so, the NQA may use careful prompt engineering to arrive at an acceptable performance as follows:

- when the first persona generates an incorrect query, the same question is asked again, but prefixed with the error message to allow the model to refine its query.
- the second persona is not provided the query results alone, but the initial question and the resulting query from the first persona.

Although two personas are considered, it should be noted that that the NQA is a single model, and its behavior depends on the prompt. In the initial demonstration (supervised) phase, personas cannot be trained end-to-end as SQL queries are, of course, non-differentiable. However, they can be trained in a collaborative manner: for a given demonstration (i.e., pair of question/answer), the model is first asked to produce a SQL query based on the question, then an answer based on the query outcome, and the resulting loss is backpropagated twice. The initial demonstration phase is then followed by training a reward model, and a reinforcement learning phase.

In various embodiments, the techniques herein also introduce a mechanism whereby the NAAI (agent 402) and the NQA (agent 404) are jointly fine-tuned to operate together. More specifically, the NAAI may be modified to operate in two steps:

- 1. Given an input (either as persona “User-facing” or “Operator-facing”), it is first allowed to generate a query to the NQA.
- 2. The input and the answer of the NQA are then passed to the NAAI to generate the final output.

Again, this procedure being non-differentiable, the system may employ a collaborative training strategy whereby the NQA and the NAAI are fine-tuned separately on a joint loss obtained from a reward model trained by human labelers.

Optionally, in some embodiments, the NAAI may also be configured to be proactive by receiving triggers from a network assurance service. These triggers may take the form of structured alerts that resemble logs, which may cause the NAAI to engage either with the network team or the end users, depending on the situation. For instance, if a violation of the voice QoS template on the SD-WAN router of Building X is observed, and no re-routing is possible, the NAAI could proactively reach out to users that are actively using voice applications (e.g., Webex) and inform the network team that it has done so.

Similar to previous components, this capability would require some demonstration to become effective, although in-context learning (ICL) could prove quite effective here, for instance by crafting some instruction with the input:

If relevant, inform active users of application X that the QoS requirement on the WAN uplink of building X is not met.

Inform the on-call network engineers that users impacted by the following issue have been notified: <log message>.

This type of “augmented” prompt not only contains the details necessary for the NAAI to identify the users, the severity of the problem (from the log message, etc.), and the type of application, but it also explicitly instructs the model to notify active users (i.e., not all of them) of a given application, and only if relevant (i.e., accounting for their preference, which they might have expressed in previous conversations).

FIG. 5 illustrates an example 500 of conversational network assurance being performed in a network, according to various embodiments. As shown, assume that end users