EP4681032A1

EP4681032A1 - A controller for configuring an operating mode of a first controller and a method thereof

Info

Publication number: EP4681032A1
Application number: EP24709432.9A
Authority: EP
Inventors: Peter Deixler; Muhammad Mohsin SIRAJ
Original assignee: Signify Holding BV
Current assignee: Signify Holding BV
Priority date: 2023-03-16
Filing date: 2024-03-11
Publication date: 2026-01-21
Also published as: WO2024188926A1; CN120813909A

Abstract

A method of configuring an operating mode of a first controller, wherein the first controller is arranged for interacting with a second controller and each controller is arranged for controlling an actuation and/or a sensing system in an environment, wherein the operating mode comprises an explainable mode and a non-explainable mode, and the first controller in the explainable mode is arranged for transmitting explainability information to the second controller related to control decisions of the first controller, and wherein the method comprises receiving, at the first controller, a request for an interaction from the second controller and switching the operating mode of the first controller based on a characteristic of the requested interaction with the second controller and/or a characteristic of the second controller.

Description

A controller for configuring an operating mode of a first controller and a method thereof

FIELD OF THE INVENTION

The invention relates to a method for configuring an operating mode of a first controller. The invention further relates to a controller, a system, and a computer program product for configuring an operating mode of a first controller.

BACKGROUND

Connected lighting refers to a system of one or more lighting devices which are controlled not by (or not only by) a traditional wired, electrical on-off or dimmer circuit, but rather by using a data communications protocol via a wired or more often wireless connection, e.g., a wired, powerline communication, LiFi, or wireless network. These connected lighting networks form what is commonly known as Internet of Things (loT) or more specifically Internet of Lighting (loL). Typically, the lighting devices, or even individual lamps within a lighting device, may be equipped with a wireless receiver or transceiver for receiving lighting control commands from a lighting control device according to a wireless networking protocol such as Zigbee, Wi-Fi or Bluetooth.

Connected lighting system can be a part of a smart home system which comprises other connected or smart systems arranged for monitoring and/or controlling home attributes such as climate, entertainment systems, and appliances. It may also include home security systems such as access control and alarm systems. When connected with the Internet, smart (home) devices are an important constituent of the Internet of Things (loT). Although the convention of smart ‘home’ system is used, the concept of connected/smart systems are equally used in other non-home environments such as offices, retail, hospitals, public spaces such as parks and squares etc.

SUMMARY OF THE INVENTION

The inventors have realized that the interactions between different smart systems in a multi-smart systems environment are constantly increasing and evolving. Due to these ever-increasing interactions, the determination of control decisions by a smart system is not only based on the individual monitoring/criteria of the smart system but is also influenced by the interaction with other smart systems in the environment. Since, these smart systems are aimed at improving user experience in the environment, the determination of control decisions to provide optimal user experience in a multi-smart systems environment becomes challenging.

It is therefore an object of the present invention to optimize the determination of control decisions of smart systems in view of the interaction with other smart systems in a multi-smart systems environment. The determination may comprise inferring or predicting control decisions from a machine learning model based on individual monitoring/criteria and further based on interactions with other smart systems.

According to a first aspect, the object is achieved by a method of configuring an operating mode of a first controller, wherein the first controller is arranged for interacting with a second controller and each controller is arranged for controlling an actuation and/or a sensing system in an environment, wherein the operating mode comprises an explainable mode and a non-explainable mode, and the first controller in the explainable mode is arranged for transmitting explainability information to the second controller related to control decisions of the first controller, and wherein the method comprises receiving, at the first controller, a request for an interaction and switching the operating mode of the first controller based on a characteristic of the requested interaction with the second controller and/or a characteristic of the second controller.

The method relates to switching or configuring an operating mode of a first controller which is arranged for interacting with a second controller. The first and the second controller may be arranged for controlling an actuation and/or a sensing system. The actuation and/or sensing system may comprise attributes such as lighting, climate, automatic windows, entertainment systems (audio, beamers, TV), and appliances, e.g., HVAC system, connected lighting system, audio/video system, scent dispenser, windows blinds etc. The first and the second controller with their respective actuation and/or sensing system may each form a smart system. A smart system may comprise functions of sensing, actuation and/or control in order to describe and analyze a situation in an environment. A smart system may be further arranged for determining control decisions based on the available data in a predictive or adaptive manner, e.g., by using machine learning models, thereby performing smart actions.

The first and the second controller may comprise networking capabilities to communicate with each other and with other smart systems. For instance, the smart system comprises transceivers (or transmitter/receivers) to transmit and/or receive communi cation/ control signals to and from the other smart systems. In most cases the “smartness” of the system can be attributed to autonomous operation based on closed loop control, energy efficiency, and networking capabilities.

The environment may be an indoor environment such as a home, a hotel, an elderly care home, a hospital, an office, a grocery/shopping store etc., or an outdoor environment such as a street, a parking lot, a sports stadium etc.

The operating mode comprises an explainable mode and a non-explainable mode, and the first controller in the explainable mode is arranged for transmitting explainability information to the second controller related to control decisions of the first controller. The explainable mode may comprise an explainable inference mode. In an example, the second controller may also be arranged for, for instance in the explainable mode, transmitting explainability information to the first controller. The first and the second controller may comprise or be arranged for using machine learning models to determine at least some of the control decisions for controlling the respective actuation and/or sensing system. A machine learning model may comprise a mathematical function or representation of a relationship between input(s) and output(s). A model is the result of a machine learning algorithm applied to a training data set. A model is often a parametrized mathematical formula, where parameters are learned by a machine learning algorithm. Given input data, a model can produce a classification label or a regression value directly, or it can produce a probability for each possible value (input).

The explainability information may comprise information related to the determination (inference) of control decisions by the machine learning model, e.g., how machine learning model works and/or why a particular control decision is determined by the machine learning model. The explainability information may comprise transparency and/or interpretability information related to the determination of control decisions and of the machine learning models. The explainability information provides understanding of the machine learning models and the control decisions determined by the models. For example, the information may comprise the extent to which a cause and effect can be observed within an actuation and/or sensing system. Or, to put it another way, it is the extent to which it can be predicted what is going to happen, given a change in input or machine learning algorithmic parameters. Additionally, or alternatively, the information may comprise the extent to which the internal mechanics of a machine learning model can be explained in human-like terms. Human like terms may utilize easy-to-interpret cause and effect relationships. In other words, explainability refers to the model' s ability to unravel relationships within the model. The transfer of explainability information to the second controller may help the second controller understand why a particular control decision is taken by the machine learning model of the first controller. In an example, the explainability information needs not to be necessarily in a human-like terms since the information is shared with the second controller and not directly with a user. The information may be shared to the user via the second controller. Alternatively, the information may be shared with both the second controller and to a user. The human user may subsequently be provided with explainability information about the actions of the smart home system by the first and/or second controller.

The explainable and non-explainable operating mode has a direct impact on the accuracy and complexity of the machine learning model used by the first controller. For example, the non-explainable operating mode may provide higher accuracy at the expense of higher model complexity and performance with a higher need for processing power compared to explainable operating mode which may have lower model complexity to enable explainability of its inferences but suffers from lower inference performance and possibly requires less processing power. Therefore, the selection of the operating mode is an important factor in determining control decisions by the machine learning model of the first controller.

The method comprises receiving, at the first controller, a request for an interaction with the second controller. The request may be received from the second controller and/or (directly) from the user. The user may use the second controller to send the request or use a user’s device such as mobile phone, tablet etc. to send the request. The interaction may comprise a request to adapt the actuation and/or sensing controlled by the first controller. In another example, the interaction may comprise a request to provide information about the control decisions of the first controller. Yet in another example, the interaction may comprise a request to provide a joint actuation and/or sensing experience. It is to be noted that all these examples are not exhaustive (further examples are not excluded) and can be combined.

Since the method further comprises switching the operating mode of the first controller based on a characteristic of the requested interaction with the second controller and/or a characteristic of the second controller, the characteristics of the request and/or the second controller is taken into consideration for switching the operating mode, thus optimizing the determination of the control decisions of the first controller (first smart system) in view of the interaction with the second controller (second smart system) in a multi-smart systems environment. The switching of the operating mode may be realized either by (A) applying a different machine learning model e.g., a simple, highly understandable decision tree model vs non-understandable complex deep neural network or (B) the same model type is still used but being re-configured for improved explainability (e.g., drastically reducing the number of hidden layers in a neural network, decreasing the number of internal nodes in a decision tree or in in a random forest model decreasing the number of decision trees on various subsets of the given dataset). Additionally, or alternatively a hybrid of A and B approaches may be used.

The first controller may be arranged for inferring a future interaction with the second controller based on one or more of time of the day, a predetermined routine, historical data related to past interactions; and wherein the method may further comprise switching the operating mode of the first controller based on a characteristic of the inferred future interaction with the second controller.

The first controller may infer a future interaction with the second controller. For example, based on one or more of time of the day, a predetermined routine, historical data related to past interactions, the presence of the second controller' s host device, the first controller infers if it needs to collaborate with the second controller. The first controller may then assess if it is possible to achieve its current control objectives (e.g., resulting in an actuation of lighting/heating/audio/scent in the environment) with the explainable mode (e.g., using white-box machine learning algorithm) that is interpretable in itself by the second controller active in the environment. With such an inference and switching, the determination of the control decisions of the first controller (first smart system) in view of the interaction with the second controller (second smart system) in a multi-smart systems environment is further optimized.

The second controller may be arranged for controlling a sensing system, and wherein the method may further comprise switching the operating mode of the first controller based on sensing type and/or sensing outcome (and/or algorithm-related power consumption) of the sensing system.

The sensing system may comprise presence sensors for detecting the presence of persons, light sensors, humidity sensors, air quality sensor (CO, pollutants, etc.), a motion sensor, occupancy sensor (infrared (IR), passive infrared (PIR), ultrasonic, etc.), thermal sensor, an electromagnetic sensor (e.g., radiofrequency -based sensing, RADAR sensing), a structured light sensor (e.g. ToF), LiDAR sensor, an acoustic sensor, air quality sensor (CO, pollutants, etc.), video (security camera, etc.), audio (microphone, etc.) etc. Since different sensing systems may require different forms of interaction with the first controller, e.g., a vision sensing system controlled by the second controller may require explainability information from a lighting system controlled by the first controller, therefore switching of operating modes based on the sensing type and/or sensing outcome of the sensing system may further optimize determination of the control decisions of the first controller (first smart system) in view of the interaction with the second controller (second smart system) in a multi-smart systems environment.

The method may further comprise assigning a sensing priority value to the sensing outcome, and wherein the method may further comprise switching the operating mode of the first controller based on the assigned sensing priority value.

For example, the second controller may be arranged for controlling a sensing system related to sensing a health-related sensing of an elderly, e.g., fall detection, breathing detection, quality of sleep (which is known to directly affect the risk of fall of the elderly) etc. Such a sensing may be assigned a higher sensing priority value compared to sensing e.g., ambient light in the environment. Even for the same sensing type, the sensing outcome may be assigned different sensing priority values, e.g., a fall detection event may be assigned a higher priority value compared to no fall, or near fall event. Since the method may further comprise switching the operating mode of the first controller based on the assigned sensing priority value, the determination of the control decisions of the first controller (first smart system) in view of the interaction with the second controller (second smart system) in a multi-smart systems environment is further optimized.

The method may further comprise switching the operating mode of the first controller based on the request for interaction, and preferably on an assigned request priority value.

The method may further comprise assigning request priority to the received request. For example, based on time of the day, the request of a sensing task may have higher priority compared to other tasks. In another example, based on the current monitoring of temperature measurement, the HVAC actuation may have higher priority. The request based on interaction of such events may be assigned higher priority, and therefore the determination of the control decisions of the first controller (first smart system) in view of the interaction with the second controller (second smart system) in a multi-smart systems environment is further optimized.

The characteristic of the second controller may comprise a trustworthiness value indicative of a level of trustworthiness of the second controller, and wherein the method may further comprise switching the operating mode of the first controller based on the trustworthiness value.

The trustworthiness value may be determined at least based on one or more of a previous interaction with the first controller, a security certificate of the second controller etc. For example, transmitting the explainability information may be more secure to a trustworthy second controller compared to a controller which is dubious or not known to the first controller or does not have a wide-spread market adoption. Optionally, the first controller may inquire via a user interface the user about his perceived trustworthiness score of the second controller.

The characteristic of the second controller may comprise control decisions of the second controller, and wherein the method may further comprise switching the operating mode of the first controller based on control decisions of the second controller.

In an example, the second controller is arranged for controlling an actuation system and uses a machine learning model for determining control decisions, e.g., setting windows blinds, controlling climate, providing audio effects in an environment etc. In such situation the switching the operating mode of the first controller may be advantageously based on control decisions of the second controller.

The method may further comprise assigning a privacy sensitive control decision value and/or a discrimination sensitive control decision value indicative of a level of privacy sensitive and/or a discrimination sensitive control decisions of the second controller respectively, and wherein the switching of the operating mode of the first controller may be further based on the privacy sensitive control decision value and/or the discrimination sensitive control decision value.

The method may further comprise assigning a safety-critical control decision value indicative of a level of safety-critical control decisions, and wherein the switching of the operating mode of the first controller may be further based on the safety-critical control decision value.

The switching may be further advantageously based on the privacy related, discrimination sensitive related and/or safety-critical control decisions, and therefore by considering such criteria in switching the operating mode of the first controller, the determination of the control decisions of the first controller (first smart system) in view of the interaction with the second controller (second smart system) in a multi-smart systems environment is further optimized. The method may further comprise switching of the operation mode of the first controller based on an environment related contextual information.

The switching of the operation mode may be further based on the contextual information related to the environment of the first and/or the second controller. For example, the contextual information may comprise type of the environment, such as indoor environment, outdoor environment, objects in the environment, interaction between objects in the environment, interaction between an object and a person in the environment if the second controller with the actuation/sensing system is placed in a particular zone of the environment, size of the environment, occupancy of an environment etc.

In the explainable mode the first controller may use a white or a glass box model for control decisions, and in the non-explainable mode the first controller may use a black-box model for control decisions.

The first controller in an explainable and non-explainable mode may use two different types of models. For example, in the explainable mode, the first controller uses a white and/or a glass box model for determining control decisions. A white box is a model whose inner logic, workings and programming steps are transparent and therefore its decision-making process is interpretable or explainable. Simple decision trees are the most common example of white box models while other examples are linear regression models, Bayesian Networks and Fuzzy Cognitive Maps. In a glass box model all parameters and how the model comes to its conclusion are known, giving a full transparency or explainability to the control decisions of the glass box model. On the contrary, a black-box model is a model which produces control decisions without revealing any information about its internal workings.

Alternative to using different models in the explainable and non-explainable mode, the same model type may be used but the model is re-configured for improved explainability (e.g., drastically reducing the number of hidden layers in a neural network, decreasing the number of internal nodes in a decision tree).

The first controller may be arranged for controlling a connected lighting system, and the second controller may be arranged for controlling a smart speaker.

According to a second aspect, the object is achieved by a system controller for configuring an operating mode of a first controller, wherein the first controller is arranged for interacting with a second controller and each controller is arranged for controlling an actuation and/or a sensing system in an environment, wherein the system controller comprises a processor arranged for executing the steps (or at least control the execution) of the method according to the first aspect. In an example, the first controller is the system controller, and therefore, in this example the notion of system controller can be interchangeably used for the first controller.

According to a third aspect, the object is achieved by a system for configuring an operating mode of a first controller, wherein the first controller is arranged for interacting with a second controller and each controller is arranged for controlling an actuation and/or a sensing system in an environment, wherein the system comprises at least the first controller, and a system controller according to the second aspect. In an example, the system further comprises at least one lighting device.

According to a fourth aspect, the object is achieved by a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method according to the first aspect. In an example, the object is achieved by a computer program product comprising instructions which, when the program is executed by a processor of the first controller or of the system controller, cause the processor of the first controller or of the system controller to carry out the steps of the method according to the first aspect.

It should be understood that the computer program product, the system controller, and the system may have similar and/or identical embodiments and advantages as the above-mentioned methods.

BRIEF DESCRIPTION OF THE DRAWINGS

The above, as well as additional objects, features and advantages of the disclosed systems, devices and methods will be better understood through the following illustrative and non-limiting detailed description of embodiments of systems, devices and methods, with reference to the appended drawings, in which:

Fig. 1 shows schematically and exemplary an embodiment of a system for configuring an operating mode of a first controller,

Fig. 2 shows schematically and exemplary an embodiment of a system controller for configuring an operating mode of a first controller,

Fig. 3 shows schematically and exemplary a flowchart illustrating an embodiment of a method for configuring an operating mode of a first controller.

All the figures are schematic, not necessarily to scale, and generally only show parts which are necessary in order to elucidate the invention, wherein other parts may be omitted or merely suggested. DETAILED DESCRIPTION OF EMBODIMENTS

Fig. 1 shows schematically and exemplary an embodiment of a system 100 for configuring an operating mode of a first controller 120 wherein the first controller 120 is arranged for interacting with a second controller 140a-b, 150a-b and each controller is arranged for controlling an actuation 120a-e and/or a sensing system 140a-b, 150a-b in an environment 101. In this example, the environment 101 is an indoor environment i.e., a home. Alternatively, the indoor environment may be a hospital, a retail, an office etc. The environment 101 may be an outdoor environment such as a street, a parking lot etc. The environment 101 comprises different subzones, e.g., rooms, such as a living room 110a, a bedroom 110b, and a shower 110c.

In this exemplary figure, the first controller 120 is a lighting controller arranged for controlling a lighting actuation system comprising lighting devices 120a-e. A lighting device 120a-e is a device or structure arranged to emit light suitable for illuminating an environment 101, providing or substantially contributing to the illumination on a scale adequate for that purpose. A lighting device 120a-e comprises at least one light source or lamp, such as an LED-based lamp, gas-discharge lamp or filament bulb, etc., plus (or optionally) any associated support, casing or other such housing. Each of the lighting devices 120a-e may take any of a variety of forms, e.g., a ceiling mounted luminaire, a wall-mounted luminaire, a wall washer, or a free-standing luminaire (and the luminaires need not necessarily all be of the same type). In this exemplary figure, the lighting devices 120a-d are ceiling luminaires (e.g., in the bedroom 110b and in the living room 110a), and the lighting device 120e is a standing lamp. Any number and types of the lighting devices 120a-e may be present in the environment 101.

Additionally, or alternatively, the first controller 120 may be further arranged for controlling an actuation system such as windows blinds, HVAC or any other actuation system in the environment 101. Yet additionally, or alternatively, the first controller 120 may be further arranged for controlling a sensing system, e.g., temperature sensing, radiofrequency-based sensing (RF sensing), RADAR sensing etc. In an example, radiofrequency signals communicated between lighting devices 120a-e may be used for radiofrequency-based sensing for presence, motion and for other sensing applications. In another example, RADAR sensors may be integrated in the lighting devices 120a-e for sensing presence of the user 130 in the environment 101, motion of the user 130 in the environment 101 and for other sensing applications. In an example, a user device 132 such as a mobile phone, a tablet etc. may be a part of the sensing system. The other sensing applications may be related to healthcare applications such as fall detection, vital sign monitoring, sleep monitoring etc. In an example, the actuation of the lighting devices 120a-e may be based on the monitored sensing tasks.

In the exemplary figure, the second controller 140a-b, 150a-b may be arranged for controlling a vision system 140a-b (e.g., a camera in the living room 110a, and another camera in the bedroom 110b), and a smart speaker 150a-b (e.g., a speaker in the living room 110a, and another speaker in the bedroom 110b). Any other actuation and/or sensing systems are not excluded, and the second controller 140a-b, 150a-b may be arranged for controlling the other (not disclosed) actuation and/or sensing system. In the figure, the second controller 140a-b, 150a-b is shown to be integrated in the actuation/sensing system, the second controller 140a-b, 150a-b may be external to the actuation/sensing system.

The first controller 120 (and the second controller 140a-b, 150a-b) may comprise or be arranged for using machine learning model to determine at least some of the control decisions for controlling the actuation and/or sensing system, e.g., the lighting devices 120a-e. Examples of machine learning model may comprise and range from linear regression models (a line), state-vector-machines (SVM) to complex non-linear models such as deep neural networks. Machine learning models may be trained based on supervised learning, unsupervised learning, reinforcement learning etc. The machine learning models are usually complex and black box in nature. In fact, they are so complex that in many cases it is not clear how these machine learning models reach to their control decisions. For example, neural networks can be very complex and provides black-box inferences. Given an input to the machine learning model, the model does a complex calculation, and comes to a decision. However, it is not clear how the machine learning model has come to the control decision.

For a multi-smart systems environment, where different smart systems such as the lighting system 120, 120a-e, a camera system 140a-b, and a smart speaker 150a-b coexist and interact in the environment 101, understanding the control decisions of a smart system by the other smart systems is imperative. The explainability of control decisions helps smart systems to better react to a given situation in the environment 101. The explainability information provides better understanding of the machine learning model of the first controller and the control decisions determined by the model. It is further important that the first controller should have a flexibility of switching from providing explainable decisions to not providing explainable decisions or vice versa. Therefore, the first controller 120 is arranged for operating in an explainable mode and a non-explainable mode, and the first controller 120 in the explainable mode is arranged for transmitting explainability information to the second controller 140a-b, 150a-b related to the control decisions of the first controller 120. The explainability information from the first controller 120 may help the second controller 140a-b, 150a-b to better understand the control decisions from the first controller 120 and hence optimize co-existence of both smart systems.

One downside of having explainability information, is that there’s usually a tradeoff between explainability and accuracy. In general, simpler machine learning models tend to be more explainable than more complex ones. It is known in the art of machine learning that the one of the most convenient ways to achieve explainable results is to stick with intrinsically interpretable models such as Linear Regression, Logistic Regression, and Decision Trees by avoiding the use of “black box” models such as PCA, random forests or neural nets. However, usually, this natural explainability comes with a cost in model performance. The first controller 120 may use a Linear Regression model; in this case the predicted target consists of the weighted sum of input features. So, the weight or coefficient of the linear equation can be used as a medium of explaining the prediction to the second 140a-b, 150a-b if the total number of features used by the machine learning model is kept small. However, in cases where there are multiple correlated features, the distinct feature influence becomes indeterminable. Hence, to determine whether a certain linear regression model used by the first controller 120 constitutes as being black box or white box, it may be first assessed whether the model uses multiple correlated features. If yes, it is deemed black box even though it is a linear regression model. If Principle component analysis (PCA) is used, the model is considered black-box.

In contrast to black box models, glass box/white box models offer increased interpretability or explainability. Additionally, or alternatively to using different models, several methods exist in the art of explainable Al which explains the outputs of these blackbox models, e.g., most notably are:

LIME (Local Interpretable Model-agnostic Explanations) and

SHAP (SHapley Additive exPlanations)

LIME trains an interpretable model around the black box models predictions. In other words, LIME finds a model that replicates the prior one' s predictions while employing a visible reasoning process. SHAP on the other hand, takes a game theoretic approach by assigning each feature an importance value for a certain prediction. While interpreting black box models requires external tools, white box/glass box models are inherently interpretable. These approaches, e.g., LIME and SHAP are well-known in the art of explainable Al (XAI), and therefore are not further discussed here. It is to be noted that in the art of explainable Al, the explainability information is provided to a user 130, whereas in the context of the invention, such an explainability information may be communicated to the second controller 140a-b, 150a-b. The explainability information may be transferred to the user 130 either directly or via the second controller 140a-b, 150a-b is not excluded.

The first controller 120 may receive a request of interaction from the second controller 140a-b, 150a-b, and the operating mode of the first controller 120 may be switched from a non-explainable mode to an explainable mode based (or vice versa) based on a characteristic of the requested interaction with the second controller 140a-b, 150a-b and/or a characteristic of the second controller 140a-b, 150a-b. With the switching of the operating modes, the first controller 120 may optimize the trade-off between the accuracy of determining its control decisions and the explainability of the control decisions, thus improving overall user 130 experience.

In the exemplary figure, the second controller 140a-b, 150a-b is arranged for controlling a vision system, e.g., comprising a vision sensor such as a camera. The vision system 140a-b may further comprise actuation means such as a light source (not shown). The second controller 150a-b may further comprise a smart speaker 150a-b. Any number of smart devices, such as lighting devices 120a-b, cameras 140a-b or smart system may co-exist in the multi-smart systems environment 101.

Fig. 2 shows schematically and exemplary an embodiment of a system controller 210 for configuring an operating mode of a first controller 120. In an example, the system controller 210 is comprised in the first controller 120 or vice versa, e.g., the first controller 120 is the system controller 210, and the notation of the system controller 210 and the first controller 120 can be used interchangeably. In another example, the system controller 210 may be comprised in the second controller 140a-b, 150a-b or vice versa.

The system controller 210 may comprise an input unit 214 and an output unit 215. The input 214 and the output 215 units may be comprised in a transceiver (not shown) or input 214 may be comprised in a receiver and the output 215 is comprised in a transmitter, arranged for receiving (input unit 214) and transmitting (output unit 215) radio frequency signals or any wireless signal for communicating with the first 120 and/or the second controller 140a-b, 150a-b according to any suitable wireless communication protocol. The input 214 and the output unit 215 may be arranged for wired communication according to any suitable wired communication protocol. The system controller 210 may further comprise a memory 212 which may be arranged for storing communication IDs of the first 120, the second controller 140a-b, 150a- b, the lighting devices 120a-e, and/or of any actuation/sensing device. The system controller 210 may comprise a processor 213 arranged for executing or at least controlling the execution of the steps of the method according to the first aspect.

The system controller 210 may be implemented in a unit separate from the first controller 120, the second controller 140a-b, 150a-b, the lighting devices 120a-e, the user device 130, such as wall panel, desktop computer terminal, or even a portable terminal such as a laptop, tablet or smartphone. Alternatively, the system controller 210 may be incorporated into the same unit as the first controller 120, the second controller 140a-b, 150a- b, the user device 130, and/or the same unit as one of the lighting devices 120a-e. Further, the system controller 210 may be implemented in the environment 101 or remote from the environment 101 (e.g. on a server); and the system controller 210 may be implemented in a single unit or in the form of distributed functionality distributed amongst multiple separate units (e.g. a distributed server comprising multiple server units at one or more geographical sites, or a distributed control function distributed amongst the first controller 120, the second controller 140a-b, 150a-b, the user device 130, the lighting devices 120a-e. Furthermore, the system controller 210 may be implemented in the form of software stored on a memory (comprising one or more memory devices) and arranged for execution on a processor (comprising one or more processing units), or the system controller 210 may be implemented in the form of dedicated hardware circuitry, or configurable or reconfigurable circuitry such as a PGA or FPGA, or any combination of these.

To enable the system controller 210, for example, to receive or transmit communication signals, the communication may be implemented in by any suitable wired or wireless means such as a local (short range) RF network, e.g., a Wi-Fi, ZigBee, Bluetooth or Thread network, Power-over-Ethemet, power line or any combination of these and/or other means.

Fig. 3 shows schematically and exemplary a flowchart illustrating an embodiment of a method 300 for configuring an operating mode of a first controller 120. The first controller 120 is arranged for interacting with a second controller 140a-b, 150a-b and each controller is arranged for controlling an actuation 120a-e and/or a sensing system 140a- b in an environment 101. The operating mode comprises an explainable mode and a non- explainable mode, and the first controller 120 in the explainable mode is arranged for transmitting explainability information to the second controller 140a-b, 150a-b related to control decisions of the first controller 120.

The first 120 and the second controller 140a-b, 150a-b may comprise or arranged for using respective machine learning models to determine at least some of the control decisions for controlling the respective actuation 120 a-e and/or sensing system 140a- b, 150a-b. A machine learning model is a representation and preferably a mathematical representation or function that has been trained to recognize certain types of patterns in the training data. The model is trained over a set of data, generally known as training dataset. The training may be performed in supervised fashion or unsupervised fashion or in a lightly supervised fashion. An algorithm is provided that it can use to reason over and learn from those data. The data may comprise text data, image data, sound data etc. For example, the machine learning model may be trained on past samples of environmental variables and the respective control decisions and the trained machine learning model may be used to infer control decisions based on (unseen) environmental variables. The actuation 120a-e and/or sensing system 140a-b, 150a-b maybe controlled at least partially based on the control decisions inferred by the machine learning models.

In an example, the explainable mode may comprise “white-box machine learning models” and non-explainable mode may comprise “black-box machine learning models”. White-box models are machine models that provide results that are understandable for another smart system (e.g., the second controller 140a-b, 150a-b) in the smart home domain. Black-box models, on the other hand, are extremely hard to explain and can hardly be understood even by a sophisticated smart-home system.

However, in general, white box machine learning models perform worse at their tasks than black box models. Most often, the black box model has the highest model complexity and performance but requires also highest processing power (and hence highest power consumption and highest carbon intensity), while the system designers may choose for the white box model a lower model complexity to enable explainability of its inferences; however, the lower complexity model suffers from lower inference performance but possibly requires less processing power. If the white-box model uses a different, simpler model, it may have better latency.

On the other hand, the white box machine learning model may be also created by adding additional functionality to the black box model. In this case, the inference performance may remain similar, but the processing power and latency for making inferences may increase. For example, the machine learning model may comprise a deep neural network model, e.g., a convolutional neural network. The explainability information for the determination of the control decisions based on deep neural network model may comprise an explanation of the model structure of the deep neural network such as number of layers, parameters etc., which layer has learnt what kind of patterns of data, how input propagates, or which neurons are activated based on what kind of input feature etc.

In an example, the first controller 120 is arranged for controlling a connected lighting system 120a-e, and the second controller is arranged for controlling a smart speaker 150a-b. The explainability information for instance enables the second controller (smart speaker 150a-b) to understand the connected lighting system 120a-e cognition (both in realtime and after the fact). After receiving the explainability information, the smart speaker 150a-b will be able to learn/determine itself when it can fully trust the connected lighting system 120a-e vs. when the connected lighting system 120a-e should be less trusted or even distrusted. In addition, the proposed explainability information enables the smart speaker 150a-b to verify the connected lighting system’s 120a-e adherence to ethical, sustainability and socio-legal values. The first controller may also inquire via an UI the user about the trustworthiness of the second controller.

The method 300 comprises receiving 310, at the first controller 120, a request for an interaction from the second controller 140a-b, 150a-b. The second controller 140a-b, 150a-b may request via a wired or wireless communication channel using appropriate wired or wireless protocol. The interaction of the first 120 and the second controller 140a-b, 150a-b may be based on equal rights basis, i.e., there is no master-slave configuration. The first 120 and the second controller 140a-b, 150a-b may be from different manufacturers, vendors etc., and therefore the first 120 and the second controller 140a-b, 150a-b may use or employ different machine learning models. Furthermore, both the controllers, in general, may not have access to the machine learning models of each other.

The method 300 further comprises switching 320 the operating mode of the first controller 120 based on a characteristic of the requested interaction with the second controller 140a-b, 150a-b and/or a characteristic of the second controller 140a-b, 150a-b. In an example, the method 300 steps of receiving 310 the interaction request and switching 320 the operating mode is a dynamic process, e.g., the steps are performed dynamically over time or repeated over time.

When (dynamically) selecting & configuring the operating mode of the first controller 120, there is a trade-off between (A) explainability, transparency, interpretability on the first controller 120 actions the one hand and (B) sacrificing of the inference performance and inference latency of the first controller 120 due to the usage of an explainable model on the other hand (C) sustainability of the Al algorithm (power consumption & carbon footprint).

To optimize the determination of control decisions of the first controller 120 in a multi-smart systems environment 101, different criteria of if and when the operating mode of the first controller 120 may be used. For example, the switching 320 of the operating mode of the first controller 120 may be based on whether the request for interaction is coming from a trustworthy second controller 140a-b, 150a-b (e.g., the type of 3^rd party interaction the first controller 120 is going after determines to which operation mode the first controller 120 will switch).

The switching 320 of the operating mode of the first controller 120 may be based on whether the request for improved explainability is related to a high-stake, safety- critical decision which must be made by the 3^rd party second controller 140a-b, 150a-b but is impacted by the first controller' s 120 current lack of explainability and reduced trustworthiness. If the second controller 140a-b, 150a-b has to make a safety critical decision based upon the first controller’s 120 insights, the first controller 120 will honor the request of the 3^rd party second controller 140a-b, 150a-b that the first controller 120 switches 320 to a more transparent operation mode, i.e., explainable mode.

The switching 320 of the operating mode of the first controller 120 may be based on whether the 3^rd party second controller 140a-b, 150a-b needs to make a highly privacy sensitive or discrimination sensitive decision.

The first controller 120 may decide to operate in the explainable mode (e.g., using the white box model) and the non-explainable mode (e.g., using the black box model) at the same time in parallel (e.g., during this testing phase streaming all the raw sensor data to the cloud so that both the white box and black box models can run in parallel). The first controller 120 then logs the performance difference between the two operating modes in the smart home deployment scenario at hand (e.g. the placement of the sensors of the first controller with respect to the second controller in the user' s room). Running both the black box model and white box model in parallel may however be cost-prohibitive during normal operation due to the high costs of ingesting granular sensor data in the cloud. When the first controller 120 receives 310 an inbound request for an explainable mode, the first controller 120 may explain to the second controller 140a-b. 150a-b that asking for an explainable mode degrades the sensing performance and increases the sensing latency compared to the non- explainable mode and increases the carbon footprint of the machine learning model.

Similarly, the first controller 120 may also know based on its A/B testing those certain inferences during the explainable mode (e.g., emotion-or heart rate detection) are less reliable than other inferences (e.g., true presence- or breathing detection), especially when compared to the corresponding inferences during the non-explainable mode (e.g., from the black box machine learning model). In such a situation, for part of the control decision, the first controller 120 may use the black box model or a different white box model which does better at those specific features. The following criteria may be used:

(A) If the 3^rd party second controller 140a-b, 150a-b uses the input of the first controller 120 for important decisions affecting the user 130 (e.g., deciding between actuating the air cleaner vs opening an automatic window; or for deciding whether to alert the homeowner on a possible intruder), the black box model may be preferred.

(B) If the interaction with the 3^rd party second controller 140a-b, 150a-b involves only disclosing of non-privacy sensitive insights (e.g., true presence detection outputting only the occupied and unoccupied room state) which are only used by the 3^rd party system for nonprivacy sensitive tasks (e.g., rendering of light scene by a lighting device), the performance degradation due to using the white box model will be not acceptable (given that the net privacy gain by using the white box model for this user is rather limited).

(C) If the interactions between the two controllers involves sharing of privacy sensitive details about the user 130 and results in privacy sensitive actuations (e.g., health related alarms and prescribing sensory therapy & meditation), a white box model may be preferred as the first controller 120 can then justify its reasoning to the user 130, e.g., via the second controller 140a-b, 150a-b.

Besides sharing explainable information, the first controller 120 may also expect to receive explainable information from the second controller 140a-b, 150a-b. The first controller 120 may use these received messages as additional input for deciding whether to switch 320 its own operation mode.

However, the machine learning models of the many different 3^rd party vendors taking care of the different smart home verticals, which the first controller 120 is integrating with, may not be equal: Some of these 3^rd party systems (e.g., camera system 140a-b) may be able provide the first controller 120 with detailed explainability information for the decisions taken by their machine learning model. However, the second controller may not always provide the same amount of explainability to the first controller. In this case, the first controller 120 may notice that the willingness of the second controller 140a-b, 150a-b, e.g., smart speaker 150a-b to explain its control inference reasoning to the first controller 120 may have suddenly increased. Even if the second controller 140a-b, 150a-b, e.g., smart speaker 150a-b does not share the underlying reason for its attitude change, the first controller 120 may infer that the second controller 140a-b, 150a-b, e.g., smart speaker 150a-b may have recognized a critical situation in the environment 101 (e.g., a dangerous situation due to an aggressive person) which the second controller 140a-b, 150a-b, e.g., smart speaker 150a-b is concerned that smart home actuations such as from the lighting devices 120a-e or other actuators may further aggravate. Hence, the second controller 140a-b, 150a-b, e.g., smart speaker 150a-b tries to be maximally explainable in its decisions for the case that they will later be challenged by anyone on its decisions or the decisions the second controller indirectly triggered to be taken by the first controller.

However, another controller, e.g., the camera 140a-b may make their own requests to the first controller 120 but cannot switch to a white box model and hence cannot provide explainability messages to the first controller 120.

In addition to HW/SW capability limitations of certain 3^rd party second controller 140a-b, 150a-b, the first controller's 120 co-learning with another controller in the environment 101 may become a “one-way street”. For instance, the first controller 120 may share explainability information with the smart speaker 150a-b in the environment 101 but the smart speaker 150a-b does not support explainable mode. Nevertheless, it may be advantageous for the first controller 120 to share explainability information with the smart speaker 150a-b and thereby influence the actuation/sensing system controlled by the smart speaker 150a-b installed in the user' s 130 home to render a new set point (e.g., room temperature) which is more advantageous for achieving the first controller' s 120 sensory- experience goals for the user at the moment. On the other hands, certain subsystems in the environment 101 may even refuse to accept our one-way-street explainability information.

In this embodiment, the first controller 120 may utilize criteria to dynamically decide when to blindfold a second controller 140a-b, 150a-b (i.e., operate in the non- explainable mode) vs. when to actuate the glass box model (i.e., operate in the explainable mode) and transmit explainability information the second controller 140a-b, 150a-b. For instance, the first controller 120 may perform an experiment to find out if it starts sharing some explainability information with the second controller 140a-b, 150a-b may entice the second controller 140a-b, 150a-b to also share more data with the first controller 120 in return (e.g., the second controller 140a-b, 150a-b may also recognize that sharing explainability information with the first controller 120 is right now mutually beneficial). If the second controller 140a-b, 150a-b indeed starts also sharing explainability information, the first controller 120 may stop blindfolding the second controller 140a-b, 150a-b altogether or perform the blindfolding of the second controller 140a-b, 150a-b to a lesser degree.

In an example, the first controller 120 may determine the operating mode of its machine learning model based on the degree the first controller' s 120 current needs to understand the machine learning model actions taken by the second controller 140a-b, 150a-b in order to achieve, e.g., the first controller' s 120 goal for the sensory experience of the user 130. For instance, whenever the first controller 120 (e.g., the lighting controller) requires alignment with the audio streaming system (e.g., music streaming) controlled by the second controller 140a-b, 150a-b, it switches to an explainable mode with the second controller 140a-b, 150a-b in hope that the second controller 140a-b, 150a-b will also share the “what and why” of its audio streaming decisions as well as the non-audio actuations it has triggered in the past, what actuations it is doing right now, what actuations will be done next and perhaps even unveiling the detailed information and hypothesis these actions of the second controller 140a-b, 150a-b have been based on. The characteristics received from the second controller 140a-b, 150a-b make it possible for the first controller 120 (i) to confirm its existing knowledge (ii) to challenge its own existing knowledge (iii) allows the second controller 140a-b, 150a-b to actively challenge the existing knowledge of the first controller 120 (“I disagree with your conclusion”) and (iv) allows the first controller 120 and the second controller 140a-b, 150a-b to jointly generate new assumptions leading to a better coordinated, consistent sensory experience (light scene, audio streaming, temperature) for the user 130.

Once, the user 130 stops the audio (e.g., music) streaming accompanied by the entertainment lighting 120a-b controlled by the first controller 120 and the collaboration is no longer mutually beneficial, the first controller 120 and/or the second controller 140a-b, 150a- b may again revert to non-explainable mode.

In an example, the first controller 120 in the explainable mode may use or employ a glass-box model. In the glass-box inference, the first controller 120 uses a model which is purposefully designed as a "glass box", meaning that another controller 140a-b, 150a-b in the environment 101 may monitor the inputs and outputs of the first controller 120. The inner workings of the glass box model may be explained by the first controller 120 in detail to the other 3^rd party second controller 140a-b, 150a-b. In an example, a user dependent assignment of the explainable and not- explainable modes may be used. For instance, for a 3 -year-old kid being home alone, explainability of machine learning decisions on which piece of art to display with the beamer is not required while explainability of detecting an open window is important, while if his parents are also in the room, the machine learning decisions could be challenged by them and hence should be explainable.

The method 300 may be executed by computer program code of a computer program product when the computer program product is run on a processing unit of a computing device, such as the processor 213 of the system controller 210.

It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and that those skilled in the art will be able to design many alternative embodiments without departing from the scope of the appended claims.

In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. Use of the verb "comprise”, and its conjugations does not exclude the presence of elements or steps other than those stated in a claim. The article "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer or processing unit. In the device claim enumerating several means, several of these means may be embodied by one and the same item of hardware. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

Aspects of the invention may be implemented in a computer program product, which may be a collection of computer program instructions stored on a computer readable storage device which may be executed by a computer. The instructions of the present invention may be in any interpretable or executable code mechanism, including but not limited to scripts, interpretable programs, dynamic link libraries (DLLs) or Java classes. The instructions can be provided as complete executable programs, partial executable programs, as modifications to existing programs (e.g., updates) or extensions for existing programs (e.g. plugins). Moreover, parts of the processing of the present invention may be distributed over multiple computers or processors or even the ‘cloud’.

Storage media suitable for storing computer program instructions include all forms of nonvolatile memory, including but not limited to EPROM, EEPROM and flash memory devices, magnetic disks such as the internal and external hard disk drives, removable disks and CD-ROM disks. The computer program product may be distributed on such a storage medium, or may be offered for download through HTTP, FTP, email or through a server connected to a network such as the Internet.

Claims

CLAIMS:

1. A method of configuring an operating mode of a first controller, wherein the first controller is arranged for interacting with a second controller and each controller is arranged for controlling an actuation and/or a sensing system in an environment, wherein the operating mode comprises an explainable mode and a non- explainable mode, and the first controller in the explainable mode is arranged for transmitting explainability information to the second controller related to the determination of control decisions of the first controller by a machine learning model, and wherein the method comprises receiving, at the first controller, a request for an interaction with the second controller and switching the operating mode of the first controller based on a characteristic of the requested interaction with the second controller and/or a characteristic of the second controller.

2. The method according to claim 1, wherein the first controller is arranged for inferring a future interaction with the second controller based on one or more of time of the day, a predetermined routine, historical data related to past interactions; and wherein the method further comprises switching the operating mode of the first controller based on a characteristic of the inferred future interaction with the second controller.

3. The method according to any of the preceding claims, wherein the second controller is arranged for controlling a sensing system, and wherein the method further comprises switching the operating mode of the first controller based on sensing type and/or sensing outcome of the sensing system.

4. The method according to claim 3, wherein the method further comprises assigning a sensing priority value to the sensing outcome, and wherein the method further comprises switching the operating mode of the first controller based on the assigned sensing priority value.

5. The method according to any of the preceding claims, wherein the method further comprises assigning a request priority to the received request and switching the operating mode of the first controller based the assigned request priority value.

6. The method according to any of the preceding claims, wherein the characteristic of the second controller comprises a trustworthiness value indicative of a level of trustworthiness of the second controller, and wherein the method further comprises switching the operating mode of the first controller based on the trustworthiness value.

7. The method according to any of the preceding claims, wherein the characteristic of the second controller comprises control decisions of the second controller, and wherein the method further comprises switching the operating mode of the first controller based on control decisions of the second controller.

8. The method according to claim 7, wherein the method further comprises assigning a privacy sensitive control decision value and/or a discrimination sensitive control decision value indicative of a level of privacy sensitive and/or a discrimination sensitive control decisions of the second controller respectively, and wherein the switching of the operating mode of the first controller is further based on the privacy sensitive control decision value and/or the discrimination sensitive control decision value.

9. The method according to claim 7, wherein the method further comprises assigning a safety-critical control decision value indicative of a level of safety-critical control decisions, and wherein the switching of the operating mode of the first controller is further based on the safety-critical control decision value.

10. The method according to any of the preceding claims, wherein the method further comprises switching of the operation mode of the first controller based on an environment related contextual information.

11. The method according to any of the preceding claims, wherein in the explainable mode the first controller uses a white or a glass box model for control decisions, and in the non-explainable mode the first controller uses a black-box model for control decisions.

12. The method according to any of the preceding claims, wherein the first controller is arranged for controlling a connected lighting system, and the second controller is arranged for controlling a smart speaker and/or a smart display.

13. A system controller for configuring an operating mode of a first controller, wherein the first controller is arranged for interacting with a second controller and each controller is arranged for controlling an actuation and/or a sensing system in an environment, wherein the system controller comprises a processor arranged for executing the steps of the method according to any of the preceding claims.

14. A system for configuring an operating mode of a first controller, wherein the first controller is arranged for interacting with a second controller and each controller is arranged for controlling an actuation and/or a sensing system in an environment, wherein the system comprises: at least the first controller, a system controller according to claim 13.

15. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method of any one of claims 1-12.