Combining machine learning and humans to create interactive systems


Design an interactive installation for the upcoming Museum through the City event taking place in Eindhoven, the Netherlands. The installation will need to use personalized visitor data collected by both your design as well as the other installations present.


The AI Factory


User Testing

Visual Design

Script Writing



Machine learning is spreading quickly, partly because software to implement machine learning is now widely available (e.g., in Python). This means that machine learning is now sufficiently mature and available to be used as a design material. This project was aimed at designing with artificial intelligence in actual context. It provided awareness into different types of algorithms and their application in real-life settings.

Data collected during the user tests



Two puzzles with different difficulty levels were created to collect the necessary data. The first puzzle was a lamp which was initially created to investigate the possibility of the overall concept. The second puzzle included a hand-made radio concept which asked the tester to fit two components into their appropriate place after which they had to figure out the right way of wiring and connecting all the components so that the radio would work. Additional information was gathered on age, self-assessed technical skills and city of residence.



The goal of our data mining was to create a model that could predict how long a participant would take to complete the various puzzles we had.

Lamp puzzle - When the linear regression was trained with only the participant age available (as that is the only attribute we are sure of having) and a split of the data into training and testing set at 50%, its relative performance was unimpressive, but its absolute performance in a usable range. In an effort to get a glimpse of how the model might behave when more data is available, the other attributes were added and attribute selection was disabled. With attribute selection disabled the performance of the model fell.

Radio puzzle - Due to a changing testing set up halfway through the experiment, only five data instances could be used, virtually excluding the possibility of meaningful results. This was confirmed when a linear regression model was trained at default settings with a split of 66%, which showed it excluded all attributes and simply returned the mean value of all participants. It was clear that a model would require more training data to be useful in such a scenario. By

retraining the models after every participant the system should become better

at deciding which puzzle to present to a visitor as it is on display longer.

1. Lamp: linear regression with only participant age available and percentage split of 50%.

2. Lamp: linear regression with attribute selection disabled and percentage split of 50%.

3. Radio: linear regression with default settings and percentage split of 66%.

4. Radio: linear regression with attribute selection disabled and percentage split of 66%.


The Factory did not just tell the participant facts about Eindhoven. Instead, it offered a puzzle where the participant needed to construct a simplified version of a machine with history in Eindhoven. In doing so, we implicitly conveyed a part of the history while offering an enjoyable activity at the same time. Also, the jokes integrated into the narrative of the accompanying audio included little snippets of Eindhoven's rich history.


Our demonstrator of The Factory was a scaled-down version of what we envision its final form would be. Ideally, it would be a small building made to look like an old factory with levers, buttons, and flashing lights all over it. A muffled voice, seemingly coming from a big hall inside the building, would then shout at bystanders begging to come and rescue his failing factory by repairing his broken products.

The demonstrator still featured the core qualities of the envisioned installation, which were: a talking factory director calling for help and guiding participants through the puzzles, actuated doors which could disclose different puzzles to the participant, two puzzles of different difficulties to allow for participants of different age and capability, and a button with a flashing light to add distractions to make the puzzles harder if deemed necessary by the system. 


The Factory used artificial intelligence on multiple levels. First of all, as the puzzles were designed to cover a range of complexity, they were selected for the user via our algorithm. The choice for a specific puzzle was primarily based on the age of the participant and the data from other projects present at Museum through the City, such as technical background or interests. Depending on how well a visitor was doing, the director would speak up either with tips or tricks, exclaiming it was in pain, be angry about the pace, or come up with other extravagant remarks. 



The demonstrator had multiple branching paths, spanning two major storylines matching the two puzzles. Minor branches within those storylines provided an opportunity for further fine-tuning of the difficulty level, as well as the story arc. The audio produced by our factory, and its director, played an indispensable part in creating this engagement; no less than 33 audio fragments were recorded to create a sense of personal interaction with the factory director.

Click here for the full report