Dynamic Discrete Spaces

[quaquel]

Currently, all DiscreteSpace subclasses rely on @cache and @cached_property for neighborhood-related functionality. This is critical for performance. However, if you want to represent a dynamically changing discrete space (e.g., a dynamic network), this is currently not possible.

It seems that functools offers a clear_cache function that can be called on anything decorated with a caching decorator. This means that it is possible to retain the performance benefits of caching while also being able to clear it on specific parts of the DiscreteSpace.

Something like the following is roughly what would be needed. Note that this is not tested and just quickly put together based on some docs.

class DynamicNetwork(Network):
	
	def add_cell(cell:Cell):
		self.G.add_node(cell.coordinate)
		self._cells[cell.coordinate] = cell.coordinate

	def add_edge(cell1:Cell, cell2:Cell):
		self.G.add_edge(cell1.coordinate, cell2.coordinate)

		for cell in [cell1, cell2]:
			cell.get_neighborhood.cache_clear()
			cell.neigbhoorhood.cache_clear()
			cell._neighborhood.cache_clear()
			self._connect_single_cell(cell)

	def remove_cell(cell:Cell):
		neigbors = cell.neighborhood

		self.G.remove_node(cell.coordinate)
		
		# iterate over all neighbors
		for cell in neighbors:
			cell.get_neighborhood.cache_clear()
			cell.neigbhoorhood.cache_clear()
			cell._neighborhood.cache_clear()
			self._connect_single_cell(cell)


	def remove_edge(cell1:Cell, cell2:Cell):
		self.G.remove_edge(cell1.coordinate, cell2.coordinate)

		for cell in [cell1, cell2]:
			cell.get_neighborhood.cache_clear()
			cell.neigbhoorhood.cache_clear()
			cell._neighborhood.cache_clear()
			self._connect_single_cell(cell)

[Sahil-Chhoker]

Can I ask how does caching prevent the dynamic nature of these spaces?

[quaquel]

The neighborhood of a given cell is cached. If you next change the neigborhood by either removing or adding a cell to it, you have invalidated the cache.

[Sahil-Chhoker]

Does the cache not update with the new neighborhood?

[quaquel]

no caching takes whatever was the return of the first ever time the function was called.

[Sahil-Chhoker]

Would it be beneficial in any way that we build our own caching function that automatically updates according to the necessary neighbor or is it too time consuming (performance wise)?

[quaquel]

Why implement something ourselves that is well covered by an existing library? Also, how would any cache know it needs to be updated? The cache_clear function is provided explicitly to cover our use case.

Separately, thinking about the issue a bit more, my hunch is to implement the add_cell, remove_cell, add_edge and remove_edge (probably these need to be renamed to something a bit more generic) at the level of DiscreteSpace. In this way, all subclasses are automatically dynamic.

For example, a while ago someone wanted to add walls to a space. With a remove_edge method this becomes trivial to do.

[Sahil-Chhoker]

Thanks for this discussion, I think I kind of get it now!

[quaquel]

I opened a PR: #2755 so let's move further discussion to there.