{"id":703,"date":"2024-09-11T16:55:24","date_gmt":"2024-09-11T16:55:24","guid":{"rendered":"https:\/\/inventorics.com\/?p=703"},"modified":"2024-09-11T17:03:17","modified_gmt":"2024-09-11T17:03:17","slug":"clustering-of-an-inventory","status":"publish","type":"post","link":"https:\/\/inventorics.com\/?p=703","title":{"rendered":"clustering of an inventory"},"content":{"rendered":"\n

Applying a clustering algorithm to an inventory of irregular unique objects can help to reduce the complexity involved in designing with such parts significantly. By dividing the inventory items into groups with similar characteristics, each group can then be represented by one “proto-part” instead, therefore reducing the amount of unique elements to be handled in setting up aggregation logics and the aggregation processes.
The decision about the number of different groups (Fig. 1)<\/em> can be completely left to an algorithm (depending on various predefined – by the programmer – conditions) or be manually determined by the user\/designer.<\/p>\n\n\n

\n
\"\"
Fig. 1: clustering of inventory with different amounts of groups (“proto-parts”)<\/em><\/sub><\/figcaption><\/figure>\n<\/div>\n\n\n\n\n\n\n

Furthermore, the reduction to a certain amount of groups allows training of an AI-model<\/em> (future blog post)<\/s><\/em>, that will be able to continually produce aggregations from the representing “proto-parts” (Figs. 2 + 3)<\/em>; instead of the need of a new training phase for each new set of parts – as long as each new part fits in one of the initial groups of the training set.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n
\n
\"\"
Fig. 2 and 3: two topologically equal aggregations with different arrangements of “proto-parts” representing all the available parts from a possible inventory<\/sub><\/em><\/figcaption><\/figure>\n<\/div>\n\n\n

Especially in an industrial product lifecycle<\/a>1<\/a><\/sup> this offers a lot of potential as most products’ geometric variance and structural potential tends to fall within a very limited range, due to norms and regulations and fabrication (mass production) constraints. For industrial artifacts it is much easier to find a smaller\/limited representative set of “proto-parts” with lesser variance of the individual parts to their “proto-part” than for example in natural materials like tree branches (Fig. 4)<\/em> or crushed rocks.<\/p>\n\n\n

\n
\"\"
Fig. 4: A set of 42 tree branch forks grouped into 5 clusters. A much larger variance between the individual items within the same cluster are visible, especially when compared to industrial objects like e.g. bike-frames (see Fig.1)<\/em><\/sub><\/figcaption><\/figure>\n<\/div>\n\n\n

In the course of the preceeding research project Conceptual Joining<\/a> we assembled spacial frameworks made of beechwood branch forks and used one single “proto-branch” for the initial set up which then got populated with the actual items from the sets of branch forks. This led to significantly different grid spacings between the different prototypes (Fig. 5)<\/em>.<\/p>\n\n\n

\n
\"\"
Fig. 5: The topologically equal grids from two different “proto-parts”. The proto-part used in the left hand grid is generated from a set of branch forks with much larger crotch angles than the right hand one.<\/sub><\/em> (www.conceptual-joining.com)<\/sub><\/em><\/figcaption><\/figure>\n<\/div>\n\n\n

Replacing the proto-parts with the geometry of the actual parts will then change the global form of the structure. That means the form of the initial aggregation will change during the replacement with the actual parts <\/a>depending on how close the approximated proto-part is or on how many different proto-parts are used for its generation.<\/p>\n\n\n\n

Having the ability to adjust the amount of representative proto-parts to design with provides flexibility and designerly freedom to choose from the complete range between a set of unique elements to one single proto-part. The complexity of setting up aggregation logics can be adjusted in relation to i.a. the type of artifacts in use or the skill-level of the designer.<\/p>\n\n\n\n

K-means <\/strong>Clustering<\/h2>\n\n\n
\n
\"\"\/<\/a>
Fig. 6: “K-means uses an iterative refinement method to produce its final clustering based on the number of clusters defined by the user (represented by the variable K) and the dataset. It tries to make the data points of same cluster as similar as possible while also keeping the clusters as different as possible.<\/sub><\/em>“2<\/a><\/sup><\/figcaption><\/figure>\n<\/div>\n\n\n

Based on an implementation by Andrea Rossi<\/a> of the K-means clustering method<\/a> from the previous project Conceptual Joining<\/a> the first step was to implement the K-means++ initialization, a smarter initialization method, which spreads out the initial cluster centers more evenly.<\/p>\n\n\n\n

Regular K-means Initialization<\/h3>\n\n\n\n
    \n
  1. Random Initialization<\/strong>:\n