Possible short term performance improvements:

• pathfinding
  • path contraction to POIs/line nodes--add nodes along the same line as node neighbors, or add major transfer points (this has previously reduced performance)
  • multithreading/hardware utilization (likely not necessary)
• citizen update loop
  • delay updates/checks on citizens that aren't expected to get to their destination for a while (some sort of priority queue/other data structure)
  • path contraction post-generation (this has previously created issues and failed to improve performance, but does reduce memory usage significantly)
  • explore options for vectorization/individual update optimization (likely not helpful)

Possibilities for extension:

• custom and accurate train counts, speeds (from https://new.mta.info/schedules, The Weekender)
• tweaks to pathfinding algorithm for better realism
  • dynamically update weights for algorithm using train ETAs (incorporate headways)
    • use D*?
  • adjust algorithm hyperparameters (and add variance to individual agent heuristics) for realistic choice-making
• drawing "complex lines" instead of straight lines between stations (I already have the data)
  • making trains follow those lines
  • giving trains basic physics (accel, decel, turning)
    • or at least precompute this for realistic headways
• randomized delays and train backups/queues
• schedule switching (weekend/late-night/etc.)
• incorporate other transit modes (e.g. bus)