[Python] Approach for implementing neural network that takes lists of varying sizes as input

I'm working on a project that simulates stacking boxed/polygon shapes onto a pallet and is currently in the process of implementing an agent that can determine where to place the polygons and is currently thinking of using a neural network of some sort. The environment is defined as a mesh of triangles and can be summarized as follows:

• A boundary area that defines the areas in which the boxes/polygons can be placed, represented as position points.
• pallet defined as the initial boundary of the mesh.
• boxes/polygons when inserted into 'pallet' are partitioned into triangles and updates the bounding line.

provided is a simply made 2D representation of the environment. Red area represents the inserted polygons with the blue lines representing the partitioning into triangles. The green lines represents the edge of the bounding area.



My current approach to implementing the agent is to give as input the polygon that is to be inserted as a list, as well as the boundary area as a list to determine a position of the polygon. My issue is that both the list for the polygon and the boundary area vary in size for each newly inserted polygon.

Are there any existing approaches to use inputs of varying sizes or to restructure the input data?

Continue reading...