LearningHome: Cooperative and active learning for the responsible improvement of energy practices in the residential sector – LearningHome

LearningHome relies on an original combination of methods at the intersection of interactive artificial intelligence, human–computer interaction, and socio-technical analysis, specifically designed to address the complexity and variability of residential energy practices.

A first core component is interactive machine learning, which enables the system to selectively query occupants when additional information is expected to significantly improve its understanding. This approach minimizes user burden while ensuring that critical contextual information—often inaccessible through sensors alone—is captured efficiently.

Complementing this, the project develops a cooperative learning framework in which occupants are not treated as passive data providers but as active contributors to knowledge construction. Sensor data are combined with occupant annotations describing activities, intentions or perceived outcomes. The system progressively aligns its internal representations with occupants’ own categories, handling inconsistencies through iterative dialogue and adjustment.

A second key technological pillar is the use of case-based reasoning (CBR). Instead of relying on predefined physical or statistical models, the system leverages a database of past situations to generate contextualized recommendations. Similar situations are retrieved and adapted to the current context, while mechanisms are introduced to account not only for successful cases but also for failures, improving robustness and exploration capabilities.

The project also introduces a self-experimentation framework at the household level. Occupants can formulate questions about their practices (e.g., the impact of a temperature setting or appliance usage), configure experiments, and analyze results through appropriate visualizations. This transforms users into experimenters of their own environment and supports a continuous loop of inquiry, observation and adjustment.

To support these processes, LearningHome integrates multi-source data fusion, combining time-series sensor data (energy consumption, environmental variables) with qualitative inputs from occupants. This hybrid data model enables the representation of both measurable phenomena and subjective perceptions.

On the interaction side, the project develops engaging human–computer interfaces designed to foster introspection and anticipation rather than prescription. These include interactive annotation tools that help users relate observed energy patterns to their activities, as well as novel interaction modalities (e.g., tactile interfaces for thermal perception) that allow users to anticipate the consequences of their actions.

Finally, the project adopts an interdisciplinary methodological framework, combining quantitative analysis of behavioral data with qualitative insights from social sciences. This enables not only the detection of changes in practices but also the understanding of their underlying drivers.

The LearningHome project has produced a coherent set of scientific and technological results that collectively demonstrate the feasibility and relevance of an occupant-centered, cooperative approach to residential energy management.

A first major result is the empirical confirmation of the limitations of directive and model-driven approaches. Building on the outcomes of the INVOLVED project, experiments have shown that systems perceived as prescriptive or partially ignorant of occupants’ intentions tend to be rejected, even when supported by advanced prediction and explanation mechanisms. This finding highlights that acceptability depends not only on algorithmic performance but also on the role assigned to occupants within the decision process.

In response, LearningHome has formalized and validated a new paradigm of inclusive energy management, where the system acts as a facilitator of understanding rather than a controller. This paradigm is operationalized through several key contributions.

A first contribution is the development of a case-based reasoning recommendation engine adapted to residential contexts. This engine leverages past situations to generate contextualized suggestions, integrates mechanisms for exploration, and explicitly accounts for unsuccessful cases to improve robustness.

A second contribution is the design and implementation of a self-experimentation framework, enabling occupants to formulate questions, configure experiments and interpret results based on both sensor data and annotations. This approach extends traditional eco-feedback systems by transforming users into active experimenters of their own practices.

A third result concerns the development of interactive and cooperative learning mechanisms. These methods enable the system to combine quantitative sensor data with qualitative occupant inputs, progressively aligning system representations with occupants’ perceptions. This contributes to building shared, interpretable knowledge about energy practices.

On the interaction side, the project has produced innovative interface concepts, including interactive annotation tools and anticipatory interaction devices. Experimental studies have shown that these interfaces foster introspection and can trigger reflection and adjustments in user practices without relying on prescriptive strategies.

Methodologically, the project has also proposed a comprehensive framework for analyzing and detecting behavioral changes, combining quantitative indicators with qualitative explanations. Experimental results highlight the strong variability of behavioral changes across households and confirm the central role of socio-psychological factors in adoption and persistence.

In terms of technological maturity, the project has progressed from TRL3 to TRL5, with several partial prototypes validated in experimental settings.

The main distinctive feature of LearningHome lies in its paradigm shift from system-driven optimization toward occupant-centered, cooperative energy management. Rather than attempting to predict and control occupant behavior, the project positions the inhabitant as the central decision-maker and focuses on enhancing their ability to understand, interpret and shape their own practices.

This shift is operationalized through several key innovations. First, the integration of cooperative learning mechanisms enables the progressive alignment of system representations with occupants’ perceptions, acknowledging the importance of tacit and contextual knowledge that cannot be captured by sensors alone. Second, the introduction of self-experimentation at the household level transforms occupants into active experimenters, capable of formulating questions, testing practices and interpreting outcomes. Third, the use of case-based reasoning allows the system to provide contextualized and interpretable guidance without relying on complex and non-transferable physical models.

Another important feature is the emphasis on non-prescriptive and engaging human–computer interaction, designed to foster introspection and anticipation rather than persuasion or control. This contributes to improving user acceptability and long-term engagement, which are critical conditions for sustainable behavioral change.

Looking forward, several perspectives emerge. A first priority is the integration and consolidation of the different technological components (learning mechanisms, recommendation engine, self-experimentation framework and interfaces) into a unified, robust platform that can be deployed in real residential environments. This requires additional software engineering efforts to ensure reliability, scalability and usability.

A second perspective concerns the large-scale and long-term experimental validation of the approach. While partial experiments have demonstrated the relevance of the concepts, comprehensive longitudinal studies involving diverse households are needed to quantify impacts on energy consumption, comfort and behavioral dynamics.

A third direction relates to socio-economic valorization. The results suggest that future markets will likely shift from automated control solutions toward services supporting understanding, diagnosis and co-construction of energy practices. This opens opportunities for industrial partnerships, integration into existing digital platforms, and the development of start-up initiatives.

Finally, from a scientific perspective, LearningHome contributes to structuring an emerging research field at the intersection of interactive AI, human–building interaction and behavioral sciences. Future work will aim at deepening this interdisciplinary framework, particularly by improving the evaluation of behavioral change mechanisms and by extending the approach to a broader range of building services and contexts.

Although the progress in the efficiency of residential buildings, the consumption does not decrease as expected. Solution for involving inhabitants in sobriety and flexibility have already been proposed but they rely on site knowledge models. However each site is unique because of its architecture, equipment and sensors but also because of its occupants. The site models are mostly not available. LearningHome aims at developing cooperative and interactive learning for home inhabitants to confront to an Interactive Home Energy Management Aid System (IHEMAS) to yield knowledge about occupant activities and costs/comforts preferred compromise. LearningHome extends the promising concepts opened up by the ANR INVOLVED project regarding interactions by developing cooperative solutions to learn a global human-system learnt representation. The aim is to identify the practices as well as the activities of the occupants by reconciling the perceptions of the IHEMAS and its more or less numerous sensors with the perceptions of the inhabitants. These perceptions will be translated into activity labels, intentions and preferences, taking into account the volatility of the inhabitants' memory and
limited consent to interact with an IHEMAS. It induces learning methods with ad hoc notifications but also mechanisms for matching the inhabitants and IHEMAS perceptions.
It might be discrepancies between the IHEMAS and inhabitants perceptions because of a little number of sensors or because of a too high complexity in the inhabitant perceptions. These confusions must be resolved by automatically adapting for instance the generated features. Combining sensor data with labels from occupants yield a model thanks to learning algorithms. Interactive learning is a complementary method to discover the energy behavior of a site. Contrary to the INVOLVED approach, explanations and advice are generated without an a priori physical model, but by exploiting similar encountered situations. The aim is to conceive an exploratory approach guiding the inhabitants in the discovery of the effects of actions in similar situations. Inhabitants will thus be put in situation of experimenters of their environment and the IHEMAS will have the role of recording the experiments and guiding inhabitants towards new exploratory. It will engage inhabitants of residential buildings towards sober energy management through user interaction and that helps them to maintain their behavior change over time. According to J. Grudin, the future of Human Computer Interaction (HCI) are smart digital partners. Thus, the goal is to investigate mixed initiative through symmetrical co-learning interactions: both parties will inform, explain, ask, suggest and learn from the other. It is a new paradigm for IIHS, which fits well the unicity of each home where knowledge raises up from confrontation of parties. Our hypothesis is that co-learning will leverage user engagement as it puts users back in the decision loop by letting them to control the system boundaries.
LearningHome will experiment different approaches to involve occupants to be more sober and more flexible in collective residential buildings. Different behavioral levers are going to be tested. One challenge is about measuring the impact of each lever: while it is relatively easy to measure a lever impact on energy consumption, it is difficult to assess an impact on energy used for heating. The assessments of the results follow two complementary approaches: an energy performance verification protocol for measuring over a few weeks the energy impacts of levers by measuring then for extrapolating them to a year by propagating uncertainties, and the analysis of household behavior regarding energy usage.