Quvia’s Approach to Maximizing Quality of Experience Using AI & ML

By Prateek Dahale, director of engineering, Quvia

‍

Quvia, formerly known as Neuron, maximizes quality of experience (QoE) by combining artificial intelligence (AI), machine learning (ML), and rule-based systems, each tailored to meet unique customer needs and enhance user satisfaction. AI enables autonomous decision-making, adapting intelligently to real-time environmental changes, while ML refines predictions based on pattern recognition and past data. In contrast, rule-based systems operate on predefined “if-then” logic for simpler, predictable scenarios. Quvia’s adaptive framework addresses the dynamic challenges of network connectivity, mapping parameters like latency, packet loss, and bandwidth to QoE metrics for diverse applications—from video calls to streaming and browsing. 

The system evolves with continuous data, optimizing traffic routing and balancing user preferences and network performance. With Quvia’s products like Pulse, 360, Grid and Network Capacity Controller (NCC), the platform empowers customers with control over their network resources, while responding proactively to fluctuating conditions. This adaptive intelligence ensures a seamless experience, elevating customer enrichment through tailored solutions that adapt to each end-user's priorities and the ever-changing landscape of connectivity.

‍

Understanding AI, ML, and Rule-Based Systems: What Are They?‍

Before diving into how Quvia maximizes QoE for customers, we’ll break down how we think about AI, ML and rule-based systems.

1. Artificial Intelligence (AI)

• Definition: AI is the simulation of human intelligence by machines. It involves enabling machines to perform tasks that normally require human intelligence, such as reasoning, learning, problem-solving, understanding language, and recognizing patterns.
• Goal: The primary aim of AI is to create systems that can autonomously perform tasks, exhibit human-like behaviors, and improve over time.

Example: Think of AI as the brain behind voice assistants like Siri or Alexa, which understand your voice commands and respond intelligently, even learning your preferences over time to provide more personalized answers.

‍

2. Machine Learning (ML)

• Definition: ML is a subset of AI focused on developing algorithms that enable computers to learn from data and make predictions or decisions based on it. As more data is processed, the algorithm gets better at making accurate predictions.
• Goal: The objective is to create systems that can recognize patterns and improve their decision-making abilities without being explicitly programmed for every scenario.

Example: Imagine an email spam filter. Instead of manually specifying every type of spam email, ML algorithms analyze thousands of emails to learn which characteristics are common in spam messages, improving their ability to block junk emails as they encounter new data.

‍

3. Rule-Based Systems

• Definition: Rule-based systems operate on a set of predefined rules set by humans. These rules are often structured as "if-then" statements, determining how the system behaves under certain conditions.
• Limitation: While rule-based systems are straightforward and reliable in predictable environments, they struggle to adapt when scenarios become complex or unpredictable.

Example: Think of a thermostat that is set to turn on the heater if the temperature drops below 68°F. If the weather suddenly fluctuates or the temperature sensor fails, the rule-based system may not respond effectively, revealing its limitations.

‍

Quvia's Story: Maximizing Quality of Experience (QoE)

At Quvia, we focus on maximizing QoE by ensuring the best possible experience for end-users, which can include people and onboard systems (e.g., cabin or inflight entertainment systems, video telemetry, etc.), under varying network conditions. To achieve this, we first need to map network parameters like latency, packet loss, jitter, and bandwidth to QoE metrics. The process involves answering questions such as:

• Does latency (delay) affect audio calls more than packet loss (missing data)?
• What bandwidth is required to ensure a buffer-free video streaming experience?
• Is it more important for a web page to start loading quickly or to load completely in the shortest time?
• Which is more critical for user satisfaction during a video call: audio quality or video resolution?

‍

Why QoE Modeling Is Challenging

These questions reveal that the relationship between network parameters and user experience is nonlinear. For instance, a 10% packet loss isn't just twice as bad as 5%; it's significantly worse because of the compounded impact on data transmission quality. This complexity means we can't simply draw straight lines to represent these relationships.

To tackle this, we collect extensive data where users rate their experience across various traffic types, such as browsing, collaboration (e.g., video calls), streaming, and file transfers, under different conditions. These conditions include various levels of packet loss, latency, jitter, and bandwidth. With this data, we train ML algorithms to estimate QoE for any given combination of network parameters.

Introducing Quvia Pulse

The trained ML model serves as a "translator," mapping network performance metrics to actual end-user experiences. We call this capability Quvia Pulse. It enables us to take the first step in optimizing QoE by understanding how network conditions affect user experience.

‍

Beyond Monitoring: Elevating the User Experience

Knowing a user’s QoE is only the beginning. The real question is: How can we elevate the QoE and optimize it across different scenarios?

People have different priorities depending on the application:

1. File Transfers vs. Video Streaming: People are generally willing to wait a few extra seconds for a file that was estimated to take 60 seconds to download, but they’re not as forgiving if a 60-second video buffers for five seconds.
2. Connecting on a Call vs. Web Browsing: People may tolerate a five-second delay connecting to a WhatsApp call, but they would find it frustrating if a simple webpage like Google.com takes five seconds to load.
3. Audio Quality in Calls: During audio or video calls, people can handle a brief video quality drop, but they won’t tolerate a five-second audio gap, which is much more disruptive.

Tailoring QoE for Different Traffic Types

Different types of traffic (e.g., browsing, collaboration, streaming, and file transfer) don’t have uniform relationships with network parameters. For example, in collaboration services, having high audio quality is crucial, while video quality beyond a certain resolution is a nice-to-have rather than a must-have.

Quvia’s decision-making model incorporates these insights to dynamically route traffic and find the right balance between network performance and user preferences. However, for this approach to work effectively, human input is still required to customize user profiles based on their preferences.

‍

The Complexity of Traffic Routing in Our Industry

In the industries we operate in (where networks and nodes often move), optimizing network traffic is not a straightforward task. We have multiple network links available for routing data (ranging from three to 20), and we deal with 30-40 distinct traffic categories. These categories are determined by combinations of factors like:

• Traffic Type: Browsing, streaming, collaboration (e.g., video calls), and file transfers.
• Application: Different applications have unique requirements. For example, streaming video content on YouTube is different from video calls on Zoom.
• User Context: The importance of the network experience can vary depending on whether the user is a crew member, passenger, or even a particular role within a company.
• Location: The context changes based on the location within the aircraft, vessel or remote site  from where the internet is accessed.

‍

Why Is This Combination Necessary?

Providing multiple traffic categories allows customers to tailor how they want the system to behave in different scenarios, giving them greater control. It’s similar to setting the context in a conversation with ChatGPT. You tell it whether you want a brief summary or an in-depth analysis, and the response is tailored accordingly. If you don’t specify the context, the output may not fully meet your expectations. In the same way, specifying traffic categories ensures that network performance is optimized according to the priorities of different users and applications.

‍

The Challenge of Continuous Decision-Making

Even with three links and 20 traffic categories, the number of potential combinations for routing traffic is enormous, reaching billions or even trillions. Deciding which traffic should go through which network link is a continuously evolving problem, not a one-time setup. The system must make decisions in real-time, considering changes in:

• Network Performance: Parameters like latency, jitter, packet loss, and bandwidth availability fluctuate constantly.
• User Behavior: Users may switch applications, making some traffic categories more important than others at different times.
• External Factors: Weather, physical obstructions, or satellite outages can suddenly affect network performance.

‍

Why Rule-Based Systems Are Insufficient

A simple rule-based approach, where predefined "if-then" rules determine traffic routing, falls short for such a dynamic environment. Here’s why:

1. Inflexibility: Rule-based systems can only respond to situations for which rules have been explicitly defined. They cannot adapt to new or unforeseen scenarios without manual updates.
2. High Maintenance: As the network conditions change, the rules must be constantly updated, making the system unsustainable in the long run.
3. Lack of Scalability: With billions of potential combinations, coding rules for each possibility is impractical.

‍

Bridging Human Intelligence and Machine Intelligence

Human intelligence plays an essential role in enhancing our system, providing context and shaping decisions by incorporating real-world insights. By feeding in general profiles based on typical use cases, we enable the system to adapt to different scenarios. These use cases may be influenced by factors such as time of day, location, user roles, system requirements, or unexpected situations that demand flexibility. 

Let’s explore how this looks in different industries, with practical examples to illustrate the diversity of needs.

Connectivity In the World of Cruising

Life on a cruise ship involves a wide range of needs, depending on who you are and what you’re doing:

• For Crew Members: Connectivity is crucial for staying in touch with colleagues and family during long voyages. Here, collaboration tools like messaging apps and video calls may take priority over general web browsing, as seamless communication helps them perform their duties and stay connected.
• For Passengers: Leisure time is a big part of the cruise experience, including streaming their favorite shows or movies and sharing photos on social media, where buffering-free video and lag-free uploads are essential for a smooth, enjoyable experience.

Tailoring Solutions for Aviation

Aviation presents unique challenges, with passengers having different expectations depending on whether they’re traveling for business or leisure:

• Business Passengers: These travelers often need to stay productive during flights, prioritizing browsing, email access, and video calls for meetings over leisure activities. For them, uninterrupted connectivity for work is a must-have.
• Leisure Passengers: For many economy passengers, in-flight entertainment is the main focus. They may gravitate toward social media platforms like Instagram and TikTok, or prefer streaming movies and TV shows to pass the time. 

Meeting the Needs of the Shipping Industry

Shipping involves long journeys and specific communication needs:

• For Crew Members: During lengthy voyages at sea, connectivity is vital. Crew members rely on robust communication, browsing and streaming applications to stay informed, connected to family and friends, and entertained in their downtime. Our solution optimizes capacity and load-sharing to meet these diverse demands.
• On the Captain’s Deck: Priorities shift with operational needs. Reliable file transfers and web browsing are essential for accessing weather updates and navigational data to optimize voyages, ensuring both safety and efficiency. Additionally, engineering demands robust connectivity for engine and cargo monitoring, as well as preventative maintenance. Our solution guarantees the necessary upstream and downstream bandwidth for these critical systems.  

Handling Diverse Requirements in the Energy Sector

The energy industry has unique demands, especially when balancing automated processes and human oversight:

• Automated Data Services: For routine operations, the system may need to trigger data uploads to the cloud at any time of day. This could require delivery of data throughout the day or equally involve sub-surface machinery sending a multitude of real-time performance metrics, which requires efficient and timely data transfer and reliably managed traffic flow.
• Service Engineers: Performing maintenance might need reliable, high-quality video communication with an offshore team to troubleshoot and fix issues in real time. Here, video call quality becomes a priority to avoid miscommunication and ensure timely repairs.

These generalized profiles guide our algorithms to make more informed decisions. This is where human intelligence augments our system and then our models kick in.

‍

The Challenge of Continuous Decision-Making

Managing traffic routing across potentially even three different network links (supported up to 25) for 15 traffic categories is a highly complex problem. The number of possible combinations runs into billions, if not trillions, making it impractical to handle with a simple rule-based system. Moreover, hard-coding solutions for every potential anomaly is unsustainable.

‍

Using Quvia’s Decision Engine to Adapt in Real-Time

Instead of rigid rules, our system models the environment and adapts its decisions based on the current "state" or "context." The ML model we developed for Quvia Pulse acts as an interpreter, advising the system on how different network conditions will impact QoE for each traffic type. This allows for intelligent decisions, such as rerouting traffic when a satellite link goes down due to weather conditions, without manual intervention. Deciding on how to route browsing, collaboration, streaming, etc. over LEO, MEO and GEO links for best QoE.

Analogy: Think of our system as an autonomous car. It doesn’t just maintain a constant speed of 50 mph; it adjusts based on the road ahead. If there’s an obstacle, it slows down or changes lanes. If the road is clear, it speeds up to reach the destination faster. Similarly, Quvia's system continuously adapts to changing network conditions to maximize QoE, it is just not limited to a focus area of operation.

‍

Why Planning Is Essential

For many of our customers, network conditions are continually changing, and the availability of resources, like bandwidth, is no longer guaranteed. If a service has a dedicated 100 Mbps connection, planning is straightforward because the network’s capacity is predictable. However, this is not the case for our customers, as network availability and performance fluctuate constantly. Moreover, with the advent of consumption-based pricing models, where we must carefully decide how much bandwidth to use each second to avoid exceeding the allocated monthly limit, effective planning becomes even more crucial.

‍

The Two Key Elements of an AI System: Environment and Agent

An AI system typically consists of two main components:

1. The Environment: This includes everything the AI interacts with – in our case, the network and its varying conditions.
2. The Agent: This is the decision-maker that learns and acts within the environment to achieve a goal – here, optimizing QoE for users.

‍

When the Environment Can Be Perfectly Modeled

If the environment is predictable and well understood, the AI agent’s job becomes much simpler. It can follow straightforward rules or algorithms to achieve its objectives because the outcomes are easy to predict. An example of this is the Voyager 1 and 2 missions launched in the 1970s:

• Perfect Environment Modeling: The space probes were able to navigate the solar system and the Kuiper Belt with great precision. Scientists had accurately modeled the gravitational forces of planets and asteroids, enabling Voyager to calculate its trajectory years in advance.
• Minimal Algorithm Complexity: Because the environment (the movement of planets and asteroids) was well-understood, the entire navigation algorithm for the spacecraft was coded with just a few kilobytes of data. The agent (the spacecraft's onboard system) needed only simple instructions to reach its destination.

‍

The Challenge of Imperfect Environment Modeling

In contrast, our system faces the challenge of operating in an environment that cannot be perfectly modeled:

1. Unpredictable Network Conditions: Network demand can change drastically due to factors like sudden spikes in usage, equipment failures, weather conditions affecting satellite connectivity, or physical obstructions.
2. Events That Are Hard to Predict: Consider a situation where 90% of the passengers on a cruise ship start streaming the Super Bowl simultaneously. Such an event can drastically affect bandwidth availability, and it would be difficult to account for every such scenario in advance.

When the environment is this unpredictable, the agent/model has to be more intelligent. It needs to:

• Try Different Actions: The agent must make real-time decisions based on the current state of the network and learn from the results.
• Adapt Quickly: It must continuously adjust its strategy to maintain QoE across different applications and users.
• Plan for Future Demand: The agent should estimate how current actions will impact future network conditions to avoid potential issues.

‍

Learning from Chronic Locations with Problems

One aspect of planning involves identifying "chronic" locations where network issues frequently occur, such as areas with repeated connectivity problems or bandwidth limitations. The system learns these patterns over time and adjusts its resource allocation strategies:

• Saving Bandwidth for Critical Moments: In areas known for poor connectivity, the system can conserve bandwidth on alternate networks, ensuring there is enough available when the main connection degrades.
• Proactive Adaptation: By anticipating issues, the system can take preemptive actions, such as rerouting traffic or limiting less critical services to maintain QoE for more important applications.

‍

Scaling Up: Managing Thousands of Users Simultaneously

The complexity of our planning process is amplified by the scale of the network. We’re not dealing with just one or two users but potentially thousands of simultaneous users, each with different requirements and priorities:

• Real-Time Decision-Making Across Many Connections: The system has to make routing decisions on a moment-to-moment basis, considering network conditions, application priorities, and user preferences.
• Dynamic QoE Optimization: As more people join or leave the network, or as certain applications gain more bandwidth, the system must continuously optimize to maintain the best possible experience for everyone.

‍

Expanding Our Capabilities with Network Capacity Controller (NCC)

Our product called Network Capacity Controller (NCC) extends our planning and decision-making capabilities even further. It not only optimizes individual network connections but also manages resources across an entire fleet or group of nodes. Here's how NCC enhances our system:

1. Fleet-Wide Optimization
   
   • Coordinating Satellite Bandwidth: NCC determines which satellite should connect to which node and how much bandwidth should be allocated to each. This is critical when multiple ships share satellite resources.
   • Global Resource Management: By considering the needs of the entire fleet, NCC can redistribute bandwidth dynamically to ensure that all nodes maintain a satisfactory QoE.
     
     
2. Handling Cascading Effects
   
   • In one instance in our simulation, the system decided to offload bandwidth from a node in the Caribbean. This decision cascaded through multiple shared networks to fulfill a sudden increase in demand in the Mediterranean region. This kind of decision-making horizon, where a single action can ripple across regions, showcases the advanced capabilities of NCC.
   • Unheard-of Decision Horizons: Traditional network management systems don’t usually make decisions that account for such far-reaching effects. NCC’s ability to think beyond individual segments and consider global optimization represents a significant advancement in network planning.

‍

This is how Quvia is different from a rule-based system

A simple rule-based system can’t cope with the dynamic nature of our environment because it lacks the flexibility to adapt. As network conditions change, a rule-based system requires continuous updates to handle every anomaly. This is not scalable, and that results in inefficiencies that impact network optimization, performance and costs.

Quvia’s approach, which combines AI (adaptive decision-making and search algorithms) with ML, enables our system to handle both familiar scenarios and adapt to unexpected changes. This creates a truly intelligent solution for optimizing QoE in the constantly evolving landscape of network connectivity. As illustrated in the graph above, pulled from a real-world maritime deployment, the integration of Quvia's decision-making capabilities led to significant improvements in QoE.

In conclusion, Quvia's approach to maximizing QoE leverages the best of AI, ML, and customer context to deliver an adaptive, real-time solution. By understanding user priorities, modeling complex relationships, and continuously learning from new data, we provide a seamless experience for users across various applications and network conditions.

‍