Resources | LpR Article | Special Applications | Artificial Intelligence | Dec 16, 2019

AI & Lighting

LpR 73 Article, page 62: Recently, Artificial Intelligence has emerged as an evolutionary force in almost every industry, demonstrating its potential to radically change existing processes. In common literature, AI is interchangeably used with Machine Learning for which various tools have already become commonplace. Henri Juslén, D.Sc (tech.), Chief Future Illuminator, Omar Nasir, M.Sc (tech.), Data Scientist, and Javad Nouri, M.Sc, Data Scientist at Helvar Oy Ab discuss AI on the context of the lighting industry where the scope of applications of AI is similarly quite broad, impacting the various stages involved in the lighting life-cycle such as design, installation, commissioning and configuration.

The lighting industry commonly employs novel techniques in lighting design and control. Major transformative changes within the industry include the evolution of the light bulb and the introduction of inter-networked lighting components that implement protocols such as DALI. Recently, Artificial Intelligence has emerged as an evolutionary force in almost every industry, demonstrating its potential to radically change existing processes. In common literature, AI is often interchangeably used with Machine Learning for which various tools have already become commonplace. In the context of the Lighting Industry, the scope for applying AI is intriguingly broad, impacting the various stages involved in the lighting life-cycle such as design, installation, commissioning and configuration. For example, a self-learning network of lighting components can communicate and set up itself without requiring human intervention similar to auto commissioning systems used in the IT industry. This will decrease the time needed to commission new lighting installations. By observing and measuring indoor environments, an AI based lighting system can optimize and tune light parameters accordingly to impact user experience and well-being. The utility of such a system is not limited to end users or tenants but extends to other stakeholders, such as building owners and facility managers as well. A data-driven network of lighting components continuously generates data which is collected and stored at a centralized server. AI algorithms can be designed to run at the source component, such as a sensor, for decentralized, real-time decisions, or at a server for making centralized decisions. Furthermore, the collected data can be utilized for other Building Management Systems (BMS) such as Heating, Ventilation and Air Conditioning (HVAC), or access management. However, the technology is not without its caveats. Cameras augmented with AI can detect precise occupancy and movements in a room or space, but the visual feed would require strict adherence to privacy laws. Another significant challenge is the limited human understanding of AI, which impedes its speed of adoption as well. In summary, paying attention to privacy, there are a lot of opportunities for applying AI within the lighting industry with significant impact to improved user experience, comfort, productivity and ultimately profitability.


Artificial intelligence (AI) refers to the study of systems artificially built by humans that can interpret the environment they are exposed to, make inferences and take actions to maximize their successes for some pre-defined objectives.

The logic of making such decisions can be either explicitly programmed, or automatically learnt from the environment. The former is generally referred to as "rule-based" systems. An example of such a system are the so-called expert systems which operate in a specific domain. The rules of such systems are created and tuned by human experts in the domain. Expert systems for diagnosing diseases is an example of such a system which helps doctors diagnose medical conditions given the symptoms received from the patients.

Machine learning is a subdivision of Artificial Intelligence and studies the ability of systems to learn, to find patterns and to make decisions based on the data they receive from their environment without explicitly being programmed. For example, machine learning can be used to predict the chance of events happening in the future, or automatically find which items in a department store are more likely to be bought.

Figure 1 shows an Artificial Neural Network (ANNs), which are machine learning systems inspired by biological neurons. In such systems, units known as neurons work together. Every neuron receives a signal either from the environment or from other neurons and produces an output signal based on the input. Outputs of neurons can, in turn, be passed on to other neurons as input signals or can be used directly to take actions. When these networks of neurons are exposed to enough data and see the results of actions, they adapt and learn how to adjust their input so that their output results in the best outcome for the task at hand. This phase is known as training the network.

Neural networks are organized into layers, so that output of one layer becomes the input to the next layer. The number of these layers is referred to as the "depth" of the network. Generally, the ability of a network to learn from the data it is exposed to improves with the increase in its depth, which translates to improved performance for the task at hand. Deeper network, however, means more resources are needed for training the network. Artificial Neural Networks have been around since the 1940s, however, a limiting factor at the time was the computational intensity of the algorithms to train the networks.

Recently, with the advance of general-purpose computing on Graphical Processing Units (GPU) and with more data available for training the networks, deep neural networks can be trained and used more efficiently, and deep learning has been more widely known and used in the industry for real-world problems.

Although Artificial Intelligence, Machine learning, Artificial Neural Networks, and deep learning are sometimes used interchangeably it is worth distinguishing them and knowing the differences between them. Figure 2 demonstrates the relationship between these concepts in a Venn diagram.

Figure 2: Artificial intelligence vs. machine learningFigure 2: Artificial intelligence vs. machine learning

Machine learning problems fall into two categories: supervised and unsupervised learning. Supervised learning involves exposing the system to input data and its corresponding output values, whereby it learns to predict the output from the input data. An example is predicting the house prices using features such as size, neighborhood, etc. The system uses historical data on house prices and the relationship between the features of the house and the price and learns to predict the price in the future.

On the other hand, in unsupervised tasks, we do not generally know the ground truth, but we are still willing to find patterns and regularities in the data. A classic example of unsupervised machine learning task is the problem of automatic grouping of objects so that similar objects fall in the same cluster.

Supervised tasks, in turn, can be separated into Classification and Regression problems. In Regression, the output is a continuous number, such as house prices. The classification problem is defined by the ability of a system to predict the category or class of the input data. For example, predicting if the weather tomorrow is going to be sunny, cloudy, rainy, etc.


Many industries started using AI and Machine Learning long ago to deliver value. Consequently, a great deal of infrastructure and tools have been created and are at the disposal of industries that are adopting data-driven approaches in their business strategies. Hardware and software infrastructure are available and widely used in Internet of Things (IoT) for collecting data from devices in the building as well as cloud infrastructure for storing, analyzing and delivering final products from the collected data.

Cloud providers such as Amazon Web Service (AWS), Google Cloud Platform (GCP), and Microsoft Azure have the tools needed for collecting and storing the data. These cloud providers even have their own solutions for IoT tasks, such as AWS IoT, GCP IoT Core, and Azure IoT Hub.

From a software tools point of view, implementations of state-of-the-art machine learning algorithms are available and readily usable in high level languages, such as pandas and scikit-learn for data analysis and machine learning in Python, Tensorflow, PyTorch, and Microsoft Cognitive Toolkit (CNTK) for deep neural networks, etc.

This wide selection of tools and infrastructures makes it easier for new industries to adopt AI, since they do not need to start from scratch. The only thing needed is to create the pipeline for collecting the data, using available tools and to start building solutions by processing the data.

Possibilities of AI in Lighting

In general, present lighting control systems are built on the belief that we know what will happen now and, in the future, i.e. somewhere there is a mastermind (the specifier or designer) who is able to develop rules governing how lighting should work in different situations. The cruel reality is that there is no such person in any professional lighting project. The problem has been solved so far by re-commissioning or re-configuration, for which an expert is needed. Lighting systems are getting more flexible, which might mean that they are also becoming more complex to setup and require special knowledge and more resources. We therefore deal with both supervised and unsupervised machine learning problems in the lighting industry, expanding to all stages of lighting life cycle, from design to commissioning to end-user experience.

Quite often, end users either do not notice or do not understand that lighting could and should work better. This means that an expert is invited only when the problem is severe (e.g. the lights will not turn on or the lights are flickering). An expert would then have to monitor lighting, space, users and their tasks over a longer period of time to be able to ensure the optimal lighting solution at all times. Unfortunately, this is not affordable nor practical in most situations.

AI for designers

Self-learning algorithms, i.e. continuous auto-commissioning through machine learning have the potential to serve as an "expert on-site" and to help grow the adoption of controllable lighting. From the designers' point of view, this means fewer compromises. Lighting can be designed without huge tolerances, like in fade times of the systems. Future AI solutions could also help to make lighting design faster and more successful. For example, if the users of commonly used Lighting design software would store their data to a cloud database, it could then already be used to provide improved recommendations at the beginning of the project. Obviously, this is one of those examples where a complementing business model is needed to ensure a successful solution.

AI for installers

Installing a lighting system takes time, which depends on the quality of planning, building structures, lighting application, knowledge and experience of personnel, and so on. Time is equally valuable for installers as well. In the future, advanced digital twins, building information models and augmented reality supported by AI could speed up installation and decrease errors. Interestingly, a significant amount of time in the building industry is wasted due to ineffective coordination, whereas the objective is to equip the right team of technicians with the correct tools and materials at the required place and time. Solving these issues might prove to be a sweet spot for AI. However, at the moment AI can only offer indirect benefits, which are monetized more in the commissioning and configuration phase.

AI for commissioning & configuration

Helvar's ActiveAhead® solution is an example of AI being used for automatic commissioning of lights. Luminaires communicate with each other about their current light stage and learn sequential patterns in the occupancy around them. This way, they can predict the occupancy in their area using the information they get from other luminaires, thus illuminate the area even if the user is at the very edge of the lighting area. This reduces the amount of effort for commissioning and programming the lighting control, and in case there is any restructuring in the layout of the area, recommissioning is not necessary since the lights will learn and adapt to the new patterns. This example highlights the future opportunities in using AI to help commissioning and configuration, well. Using the available data, it is possible to make this part of the process easier or even fully automated. When adding more advanced sensors and cloud level processing power in the future, auto-configuration and especially continuous configuration might be the mainstream of lighting in larger buildings.

AI for end users (tenants)

The reason to design, install, commission, configure, maintain and control lighting is to make the space usable for users. Good results can be achieved without any AI, if environment and needs are not changing, by making sure that selected solutions are providing illumination well above lighting norms. Unfortunately, as the needs do often change, users of the space develop different requirements. One way to solve this is to have user interfaces that allow users to change conditions. This often leads to non-optimized conditions as users/people either don't use available user interface or ignore personal user interfaces (after an initial phase of trying them out). It has been seen that people tolerate quite bad lighting before starting to control it. Hence, control should not be left just to the users. AI can play a role here. One option is supervised learning, where the system can learn the user preferences by recording their selections. It is also possible to collect data from multiple sources and offer automatic lighting that fits user needs and make lighting recommendations.

AI for building owners and facility management

Building owners who are not tenants themselves can be incentivized by improving the profitability of their buildings. Tenants who are willing to pay substantial rents will demand that the building should be managed properly. Maintenance of technical systems is an evident use case of AI. It is already possible not only to see what the problem is and where it is, but also to predict the malfunction of a component in a system. This can be done by analyzing historical data and predicting future events. Figure 3 shows an environment where the lighting remains turned on even if there are no occupants in the space. This leads to excessive energy consumption. By combining the data of multiple sensors, fade times can be tuned according to real needs. These are examples, where maintenance and the operating cost of a building can be lower while improving the tenant's comfort in a space. The value proposition AI has to offer to building owners is the improvement in overall tenant satisfaction, which inevitably translates to greater profitability through improved rental contracts. 

Figure 3: Active lighting in empty spaces [2]Figure 3: Active lighting in empty spaces [2]

Architecture of Intelligence

Designing the architecture of a network that supports AI based decision making requires special considerations. The first step is to create a pipeline for data collection, which enables the data generating devices to connect to a server. This server is responsible for communicating with the cloud. It can be hosted either locally, in the premises of the building, or in the cloud. The most important attributes of this network are reliability, flexibility and scalability. The network must be able to continuously send data, whilst catering to any changes in its topology either in the form of rearrangement or the addition of devices.

Figure 4: Example architecture for building AI productFigure 4: Example architecture for building AI product

The next step is to collect sufficient data in the cloud and train the AI model on the dataset. The trained model can be deployed in the network using its components to host the algorithm, which should preferably be the local server or router. The advantages to this approach are availability in case of connection failure to the cloud. However, this requires the edge devices to have higher computational power, which constrains the capability of the model. The alternative is to run the model in the cloud. This allows for greater model complexity at the cost of increased communication with the device network.

Finally, the output from the trained model can be used for making decisions regarding lighting control systems. For example, the algorithm can instruct a device to turn off the lights based on the output from the AI model.

Beyond Lighting Control

Modern lighting control systems incorporate an array of sensors, including motion and thermal devices. These sensors generate data based on interactions with their users. For example, an occupant in a room will generate motion events and have a specific heat signature. Similarly, the system can also use electricity meters to determine the energy consumption of luminaires.

All these sensors measure a specific parameter, which is interpreted in the context of lighting control to make informed decisions. However, there exists an opportunity to expand the interpretation beyond lighting control systems. For example, an increase in the heat signature can indicate an increase in the number of people in a room. Similarly, variations in energy consumption patterns can provide insights into how a particular space is being utilized. These interpretations are useful as they establish a framework which allows integration of AI based lighting networks with external systems.

To put the above discussion in perspective, consider a heating and ventilation system otherwise commonly known as HVAC. If the AI is able to model the number of occupants in a room based on information from thermal and motion sensors, it can instruct the HVAC to change the air quality to improve user well-being, even in advance considering the impact delay of HVAC systems to spaces. The same information can be used to identify which rooms are more frequently used, and therefore, provide insights to a building owner on the possibility of rearranging the floor layout. By improving space utilization, a building owner can maximize profitability by renting out underused spaces. Finally, it is also possible to optimize electricity costs by analyzing energy readings from different luminaires and the pattern of people flow in different rooms.


There are various challenges related to the adoption of Artificial Intelligence in different industries. - The following sections briefly describe some of these challenges.

Human challenge

Artificial Intelligence is deeply rooted in mathematics and computer science. Recent developments related to computational and algorithmic efficiency have lowered the barriers to entry, but the overall research output is still constrained to a niche group of scientists. Therefore, it is challenging to correctly interpret the decision capability of an AI algorithm for people unfamiliar with the inner workings of the algorithm. Moreover, the concept of machine intelligence has been romanticized by the entertainment industry to the point that the general population is highly skeptical of its benefits. As a result, the most commonly cited threat vis-à-vis AI is the eventual replacement of human labor. In practice, AI is being used as a supplementary mechanism to automate banal tasks and improve decision making for user well-being. This disconnect between the interpreted and the intended purpose reduces the willingness to adopt the technology at a wider scale.

Data challenge

The first and foremost requirement of building robust AI based solutions is the availability and quality of data. The decision to incorporate such products in an existing portfolio is usually taken based on an initial analysis of available data. However, the quality of analysis is closely linked to the interpretability of data. For example, imprecisions in an audio recording device can increase accumulated noise in the output, which will reduce the effectiveness of any subsequent analysis. Similarly, data generation methodologies can be constrained by network capacity, and instead attempt to summarize the observations made by the devices. AI algorithms modelled on such abstractions will produce generalized results which might not be suitable for applications that require fine-grained control. Therefore, AI practitioners must understand the limitations of the available data, otherwise the algorithm will model imperfections in the data and will likely add reduced value to the customer.

The heightened interest in AI has had a significant impact on how companies build their future strategies. Many companies have existing software infrastructure in place that does not support data generation or collection. Moreover, it is equally difficult to extract data from legacy systems that have not been designed for this purpose. Investment in modification and upgradation of legacy systems is therefore needed to enable reliable data collection. This can impede execution of company strategies that are being developed around AI driven products. However, nascent organizations can hire consultants and data engineers when building the initial infrastructure to ensure seamless integration with AI driven products in the future.

Privacy challenge

Lastly, it is vital to talk about ethical data collection and usage. Any AI based solution that optimizes user experience will incorporate user data to a certain extent. It is important that the collected data does not violate the privacy of any individual. For example, a common application of AI is to segregate similar user profiles. Therefore, it should be ensured that any personal information cannot be used to identify individuals, but rather, the information is obfuscated before feeding it to a data analysis algorithm. There are privacy laws to ensure data misuse does not take place, but it can vary from region to region. In Europe, for example, GDPR ensures compliance with ethical data standards for any company working with user data.

The most typical example of a data privacy issue arises when using cameras to collect data. A visual feed can provide very accurate readings, such as traffic flow or the number of people in an area as shown in Figure 5. However, it also allows facial recognition making this format of data susceptible to misuse as well. Aspiring companies must consider ethical issues when specifically dealing with these data collection methods.

Figure 5: Public crowd counting systems [3]Figure 5: Public crowd counting systems [3]


Disruptive technologies have traditionally faced various challenges and obstacles in their wide spread adoption and acceptance, as is the case with Artificial Intelligence as well. However, its benefits and transformative potential are undeniable. It has had a significant impact on industries such as e-commerce and media entertainment but has made limited inroads in more conventional economic activities such as lighting and building management. Previous sections discussed some of the improvements that can be introduced in the lighting industry and the possibility of extending their utility beyond lighting as well. It further highlights associated challenges and pitfalls that AI and Lighting practitioners should bear in mind while designing future systems. This will ensure development of disruptive but reliable solutions which can positively impact user comfort and well-being, whilst maximizing profitability for various stakeholders in the lighting industry.

[1]    Wikimedia Commons,
[2]    Pixabay,
[3]    Dwivedi, Priya, 'Use a Crowd Counting AI Model for your Business', Towards Data Science, 28-08-2018,