Minimal interface.

[verri]

The way the library is implemented right now many function members are required even though simulation never call them.

I think we should discuss wheter these requirements should be kept or not.

Advantages of a cleaner interface:

• Users can setup simpler classes in ther simulation.
• If we go to the extreme and remove them (instead of just throwing not_implemented) source-code will become easier to maintain.

Disadvantages of a cleaner interface:

• If we go to the extreme and remove them, users will use much more downcast.
• (cannot think anything else...)

We can improve downcasting using compatibility function that automatically downcast them in the callbacks, I think that would be the perfect scenario.

[JorgeLuizFranco]

Checked the code here and since the function is never used, probably is the best to remove it then. But how will it cause more downcast if it is not used?

[verri]

I have updated UATReinforce to work with these changes. It seems to be a very productive direction.

With the typed version of uat::permit, few downcasts were needed.

When I first designed the library I assumed two things:

• A simulation might contain more than one type of agent (that's why typed erased uat::agent)
• An implementation of an agent would work with any kind of uat::region.

The first assumption is proven correct. And I have a few ideas to keep improving uat::agent, such as exposing the interface to avoid signature mismatching.

However, the second assumption seems to be hard to implement on the user side. The Naive agent in the paper, for instance relies on many specific characteristics of the Airspace3D class.

I will try to bump the library to C++20, and investigate whether a region-specific template version of the simulation function is better. The callbacks and actions of uat::agent will need to receive something like uat::region_view to avoid unneeded copies of the region objects, reducing a lot the number of heap allocations. Some downcasts might still be needed. A clear disadvantage here is that all agents implantation will become dependent on the specific region type.

[verri]

It seems that it is now 2x faster. I have updated UATReinforce to use this version.

For me, I consider it mature enough to merge in master.

@JorgeLuizFranco take a look at https://github.com/JorgeLuizFranco/UATReinforce/tree/minimal-interface

If you are okay with the new interface, we can move on.

[verri]

A side note: fmt and jules are not required anymore.

[verri]

Reduction from 802 loc to 494 loc. Simpler way to use. Everything still works. 2x faster (not reliable info though). Let's merge!