Spaceheater

This is my third attempt at building a battle snake (src/snakes/simple.rs is the first and still available).

See https://play.battlesnake.com/profile/general-error for recent games.

This README is mostly a TODO list of reminders for myself on how to improve the snake.

The code is a bit of a chaotic mess, but feel free to steal something from /src/logic/scoring if you want :)

TODOs

Weird choices to debug

• Paranoid snake problems:
  • e3375d15-da4b-46bc-8d20-2f6531a356a4 -> turn 386 should go right and tail chase
  • d95fb988-5a7e-4135-ad1f-51bef2e95ba5 -> turn 408 why not down?
  • 7ece10cd-ea09-4298-a4ab-8ebcd771f81e -> turn 248 why not right? -> timeout! seems like MCTS implementation has a lot of stutter
  • 88b1859a-b796-41f6-9d2b-5416047bdfaa -> turn 322 should gamble up or right

The Gauntlet

• Support games with something other than 4 snakes per game
• Re-introduce random configs during next generation evaluation

Quality of life

• Add nice naming to gauntlet generations / offspring

Profiling

Create flamegraph.svg: RUSTFLAGS='-C force-frame-pointers=y' cargo run --release --features=profiling --bin replay spaceheater3 < logs/{game_id}.json.gz

Scoring

• Try max-n scoring and tree search, at least when paranoid minimax resolves to certain death.

Bugs

• "All paths are certain death, just score this board and return" -> currently does not simulate enemies in this scenario, just moves our own snake into certain death and scores the board, leading to inaccurate scoring values.

Performance

• Stateful tree search:
  • Keep tree across turns, start from subtree
  • Start computing from the game start, not the first move
  • Move request fetches relevant subtree from background worker & submits the new root back after returning
  • Stop background worker upon receiving a game end request or 2 seconds after receiving the last move.
  • One background worker per active game, background workers have lowest priority.
• SIMD: https://doc.rust-lang.org/std/simd/index.html
• wgpu compute?

Other

• de.fixnum.org running commit 66229d95f354c1d2c99a45357d4aee5087804a04 seems to be outperforming current master bc94a137e897a2b6a17618d8b1da5e6115ea7d45 on ghost -> seems to be related to the snake switching to 8bit counters in the score struct, it doesn't feed as much -> compare scores for games where both snakes were present on evening of 12/12/22 -> might be because configs contain numtype and deserializing from strings with 16bit encoded values to 8bit values will do weird, weird things

Game logic

Unsupported game features

• stacked hazards:
  • currently limited to max 3 hazards on a tile

Artwork in README is generated with Midjourney