donjpierce / traffic Goto Github PK

Simulating even a small number of cars (~30) requires too many calculations per time step. Adding traffic lights also increased simulation time. Simulating ~20 second flows takes about an hour.

Fixing this starts with optimizing how cars search for their obstacles. The present, inefficient schema is to determine if there are any obstacles in the car's field of view by looking through a list of all of the cars in the system and comparing their locations to the linspace in the field of view. A similar schema is used for finding red traffic lights. This is monumentally poor design. A more correct framework is for all cars in the system to be sorted by their position. Then each car can look for obstacles by only needing to consider a small subset of the total number of cars. Which subset is determined by the car's position and how the list is sorted.

Other performance issues could potentially be resolved by simply de-cluttering. There is probably excess data being stored that need not be stored, such as in unnecessary usage of indexed loops and indexed boolean statements. Also in the future of this base is to use data frames instead of lists and dictionaries.

Specifically when using sim.init_random_node_start_location() there is the problem of all cars simultaneously stopping in their tracks at around ~3.400 update iterations. This stoppage issue never occurs with sim.ini_culdesac_start_location().

Source of bug unknown.

Update 06/21/20:
Now this only happens with sim.init_culdesac_start_location() and not with the random node init.