An Innsbruck-based museum contacted us to optimize the distribution of 3,299 paintings in their new storage facilities located outside the city. The storage space consists of an array of double-sided, movable metal grid walls. The museum seeks to optimize the arrangement of paintings on the walls to maximize the use of available space.
The museum established various classifications and considerations for hanging the paintings on the walls:
Classification parameters
- Use the minimum possible space required to hang the paintings.
- Some paintings must be hung together on the same grid wall, such as a series of works from the same artist, in chronological order, and/or by object type.
- Paintings larger than specific X, Y dimensions or heavier than X kg should be hung at the bottom of the grid wall for easier handling.
- Paintings with a depth greater than Z dimension should be hung externally.
- Reserve areas (considering aspect ratio) for paintings involved in exhibitions, restorations, etc.
- The distance between paintings on the wall should be related to their dimensions. Larger objects require more space to handle than smaller ones.
- Other special requirements.
The database provided consisted of several Excel workbooks and folders containing images of the paintings. The client requested a report for all grid walls, showing the optimal arrangement of paintings, including the relative location (X, Y) on the wall, and the corresponding painting code ID. The graphic information displayed in the report should also be reflected in a table, detailing the wall number and complete painting data: ID, Artist, Region, Year, Description, Dimensions, Location, etc.
walls
Strategy
The project workflow was divided into three main stages: Pre-processing (analyzing the existing database), Optimization (developing software to make intelligent decisions), and Display (generating the layout plans and tables for the report). Though the process is presented in distinct stages, the workflow was continuous, with occasional back-and-forth to incorporate new parameters into the process.
Pre-processing
In this stage, the database was analyzed to devise a strategy for classification and optimization. Different scripts were written in C# and VB in Excel to survey and explore the database, identifying gaps and errors, such as swapped width and height dimensions for certain entries. Additionally, image-processing software was developed to resize and scale the painting images since many were large files. This also allowed for early control over the final file sizes.
We worked closely with the client in a workshop-based format, organizing several working sessions until the database was consistent and robust enough for the optimization process.
Pre-Processing data stage. On the right a database example received from the client and on the left, the algorithm classify the errors and void information with different colors.
Optimization
A decision tree algorithm was used to distribute the paintings across the grid walls, supported by optimization strategies implemented via the Laga library.
A Decision Tree is a widely-used algorithm in Artificial Intelligence, particularly in Machine Learning, where it is commonly applied to supervised classification problems. It maps an input to a discrete label, such as True or False. For example, given an input picture, the model might return True if the subject is male and False if female. Decision trees are useful to Decision trees are versatile and can analyze both numerical and categorical data. Since they are not based on equations, they are classified as non-parametric models. Instead, they operate through a simple, recursive structure, which makes them very fast and efficient.
The design of a decision tree algorithm is informed by the pre-processed training data.
Decision Tree representation: The algorithm builds its structure based on the input data. The goal is to narrow the search space to find the most informative questions and answers.
The algorithm works by asking a series of questions to narrow down the search space. Each question is designed to extract the most information possible, helping to locate the optimal position for each painting on the grid wall. The goal is always to reduce the number of possibilities, and the effectiveness of each question depends on the answer to the previous one.
Decision Tree Algorithm Process:
- Select the best attribute, ideally splitting the data (grid wall) in half.
- Formulate a question.
- Follow the answer path.
- Place the painting in the correct position.
- Repeat from step 1.
xamples of optimal painting distribution.
Display
The information generated by the algorithm was compiled into a report and distributed to the team responsible for hanging the paintings, as well as the museum staff.
The report, including one of the tables with painting locations. On the right is the report assigned to the grid walls.
Results:
The company responsible for transporting and hanging the paintings completed the project ahead of schedule, with no errors found in the optimization of the 3,299 paintings. The client was satisfied with the results, and some museum staff described the project as a benchmark for future endeavors. They also mentioned that staff from other museums had visited the storage facility to learn from the developed process.
Example of one the grid walls
